├── LICENSE
├── README.md
├── graphs.png
├── main.py
├── network.py
├── noises
├── ounoise.py
└── param_noise.py
├── normalized_actions.py
├── policies
├── generative.py
└── policy.py
├── replay_memory.py
└── utils.py
/LICENSE:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 |
635 | Copyright (C)
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | Copyright (C)
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | This repo contains the code for the implementation of [Distributional Policy Optimization: An Alternative Approach for Continuous Control](https://arxiv.org/abs/1905.09855) (NeurIPS 2019). The theoretical framework is named DPO (Distributional Policy Optimization), whereas the Deep Learning approach to attaining it is named GAC (Generative Actor Critic).
2 |
3 | # How to run
4 |
5 | An example of how to run the code is provided below. The exact hyper-parameters per each domain are provided in the appendix of the paper.
6 |
7 | main.py --visualize --env-name Hopper-v2 --training_actor_samples 32 --noise normal --batch_size 128 --noise_scale 0.2 --print --num_steps 1000000 --target_policy exponential --train_frequency 2048 --replay_size 200000
8 |
9 | # Visualizing
10 |
11 | You may visualize the run by adding the flag --visualize and starting a visdom server as follows:
12 |
13 | python3.6 -m visdom.server
14 |
15 | # Requirements
16 |
17 | - mujoco - see explanation here: https://github.com/openai/mujoco-py
18 | - gym
19 | - numpy
20 | - tqdm - for tracking experiment time left
21 | - visdom - for visualization of the learning process
22 |
23 | # Performance
24 |
25 | The graphs below are taken from the paper and compare the performance of our proposed method to various baselines. The best performing method is the Autoregressive network.
26 |
27 | 
28 |
--------------------------------------------------------------------------------
/graphs.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/tesslerc/GAC/841584cce21fad69950f4b2b8f691d9d3254a2d8/graphs.png
--------------------------------------------------------------------------------
/main.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import gym
4 | import numpy as np
5 | import pickle
6 | from tqdm import trange
7 | import visdom
8 | import torch
9 |
10 | from policies.generative import Generative
11 | from policies.policy import hard_update
12 | from normalized_actions import NormalizedActions
13 | from noises.ounoise import OrnsteinUhlenbeckActionNoise, NormalActionNoise
14 | from utils import save_model, vis_plot
15 |
16 | parser = argparse.ArgumentParser()
17 | parser.add_argument('--env-name', default="HalfCheetah-v2",
18 | help='name of the environment to run')
19 | parser.add_argument('--gamma', type=float, default=0.99, metavar='G',
20 | help='discount factor for reward (default: 0.99)')
21 | parser.add_argument('--tau', type=float, default=0.01, metavar='G',
22 | help='discount factor for model (default: 0.01)')
23 | parser.add_argument('--noise', default='normal', choices=['ou', 'normal'])
24 | parser.add_argument('--noise_scale', type=float, default=0.2, metavar='G', help='(default: 0.2)')
25 | parser.add_argument('--batch_size', type=int, default=64, metavar='N', help='batch size (default: 64)')
26 | parser.add_argument('--num_epochs', type=int, default=None, metavar='N', help='number of epochs (default: None)')
27 | parser.add_argument('--num_epochs_cycles', type=int, default=20, metavar='N')
28 | parser.add_argument('--num_steps', type=int, default=1000000, metavar='N',
29 | help='number of training steps (default: 1000000)')
30 | parser.add_argument('--start_timesteps', type=int, default=10000, metavar='N')
31 | parser.add_argument('--eval_freq', type=int, default=5000, metavar='N')
32 | parser.add_argument('--eval_episodes', type=int, default=100, metavar='N')
33 | parser.add_argument('--train_frequency', type=int, default=2048, metavar='N')
34 | parser.add_argument('--replay_size', type=int, default=50000, metavar='N',
35 | help='size of replay buffer (default: 50000)')
36 | parser.add_argument('--training_actor_samples', type=int, default=16, metavar='N',
37 | help='number of times to sample from the actor for calculating the losses (default: 16)')
38 | parser.add_argument('--visualize', default=False, action='store_true')
39 | parser.add_argument('--experiment_name', default=None, type=str,
40 | help='For multiple different experiments, provide an informative experiment name')
41 | parser.add_argument('--print', default=False, action='store_true')
42 | parser.add_argument('--not_autoregressive', default=False, action='store_true')
43 | parser.add_argument('--q_normalization', type=float, default=0.01,
44 | help='Uniformly smooth the Q function in this range.')
45 | parser.add_argument('--target_policy', type=str, default='exponential', choices=['linear', 'boltzman', 'uniform', 'exponential'],
46 | help='Target policy is constructed based on this operator.')
47 | parser.add_argument('--target_policy_q', type=str, default='min', choices=['min', 'max', 'mean', 'none'],
48 | help='The Q value for each sample is determined based on this operator over the two Q networks.')
49 | parser.add_argument('--temp', type=float, default=1.0, help='Boltzman Temperature for normalizing actions')
50 |
51 | args = parser.parse_args()
52 |
53 | assert args.training_actor_samples > 0
54 |
55 | env = NormalizedActions(gym.make(args.env_name))
56 | eval_env = NormalizedActions(gym.make(args.env_name))
57 |
58 | agent = Generative(gamma=args.gamma, tau=args.tau, num_inputs=env.observation_space.shape[0],
59 | action_space=env.action_space, replay_size=args.replay_size, actor_samples=args.training_actor_samples,
60 | q_normalization=args.q_normalization, target_policy=args.target_policy,
61 | target_policy_q=args.target_policy_q, autoregressive=not args.not_autoregressive,
62 | temp=args.temp)
63 |
64 | results_dict = {'eval_rewards': [],
65 | 'value_losses': [],
66 | 'policy_losses': [],
67 | 'train_rewards': []
68 | }
69 |
70 | base_dir = os.getcwd() + '/models/' + args.env_name + '/'
71 |
72 | if args.experiment_name is not None:
73 | base_dir += args.experiment_name + '/'
74 |
75 | run_number = 0
76 | while os.path.exists(base_dir + str(run_number)):
77 | run_number += 1
78 | base_dir = base_dir + str(run_number)
79 | os.makedirs(base_dir)
80 |
81 | if args.noise == 'ou':
82 | noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(env.action_space.shape[0]),
83 | sigma=float(args.noise_scale) * np.ones(env.action_space.shape[0])
84 | )
85 | elif args.noise == 'normal':
86 | noise = NormalActionNoise(mu=np.zeros(env.action_space.shape[0]),
87 | sigma=float(args.noise_scale) * np.ones(env.action_space.shape[0])
88 | )
89 | else:
90 | noise = None
91 |
92 |
93 | def reset_noise(a_noise):
94 | if a_noise is not None:
95 | a_noise.reset()
96 |
97 |
98 | print(base_dir)
99 |
100 | state = agent.Tensor([env.reset()])
101 | episode_reward = 0
102 | agent.train()
103 |
104 | reset_noise(noise)
105 |
106 | if args.visualize:
107 | vis = visdom.Visdom(env=base_dir)
108 | else:
109 | vis = None
110 |
111 | episode_timesteps = 0
112 | for step in trange(args.num_steps):
113 | with torch.no_grad():
114 | if step % args.eval_freq == 0:
115 | eval_reward = 0
116 | for test_epoch in range(args.eval_episodes):
117 | done = False
118 | eval_state = agent.Tensor([eval_env.reset()])
119 | while not done:
120 | action = agent.select_action(eval_state)
121 |
122 | next_eval_state, reward, done, _ = eval_env.step(action.cpu().numpy()[0])
123 | eval_reward += reward
124 |
125 | next_eval_state = agent.Tensor([next_eval_state])
126 |
127 | eval_state = next_eval_state
128 | results_dict['eval_rewards'].append((step, eval_reward * 1.0 / args.eval_episodes))
129 | if args.print:
130 | try:
131 | print('env: {0}, run number: {1}, step: {2}, reward: {3}, value loss: {4}, policy loss: {5}'.format(
132 | args.env_name,
133 | run_number,
134 | results_dict['eval_rewards'][-1][0],
135 | results_dict['eval_rewards'][-1][1],
136 | results_dict['value_losses'][-1][1],
137 | results_dict['policy_losses'][-1][1]))
138 | except:
139 | pass
140 | save_model(actor=agent.actor, basedir=base_dir)
141 | with open(base_dir + '/results', 'wb') as f:
142 | pickle.dump(results_dict, f)
143 |
144 | if step < args.start_timesteps:
145 | action = torch.Tensor(env.action_space.sample()).to(agent.device).unsqueeze(0)
146 | else:
147 | action = agent.select_action(state, noise)
148 | next_state, reward, done, _ = env.step(action.cpu().numpy()[0])
149 | done_bool = False if episode_timesteps + 1 == env.env._max_episode_steps else done
150 |
151 | episode_timesteps += 1
152 | episode_reward += reward
153 |
154 | action = agent.Tensor(action)
155 | mask = agent.Tensor([not done_bool])
156 | next_state = agent.Tensor([next_state])
157 | reward = agent.Tensor([reward])
158 |
159 | agent.store_transition(state, action, mask, next_state, reward)
160 |
161 | state = next_state
162 |
163 | if done:
164 | results_dict['train_rewards'].append((step, np.mean(episode_reward)))
165 | episode_reward = 0
166 | episode_timesteps = 0
167 | state = agent.Tensor([env.reset()])
168 | reset_noise(noise)
169 |
170 | if len(agent.memory) > args.batch_size and step % args.train_frequency == 0:
171 | value_loss, policy_loss = agent.update_parameters(batch_size=args.batch_size,
172 | number_of_iterations=args.train_frequency)
173 |
174 | results_dict['value_losses'].append((step, value_loss))
175 | results_dict['policy_losses'].append((step, policy_loss))
176 |
177 | vis_plot(vis, results_dict)
178 |
179 |
180 | with open(base_dir + '/results', 'wb') as f:
181 | pickle.dump(results_dict, f)
182 | save_model(actor=agent.actor, basedir=base_dir)
183 |
184 | env.close()
185 |
--------------------------------------------------------------------------------
/network.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | import torch.nn.functional as F
4 | import numpy as np
5 |
6 |
7 | class Categorical(nn.Module):
8 | def __init__(self, num_inputs, num_outputs):
9 | super(Categorical, self).__init__()
10 | self.linear = nn.Linear(num_inputs, num_outputs)
11 |
12 | def forward(self, x):
13 | x = self.linear(x)
14 | return x
15 |
16 | def sample(self, x, deterministic=False):
17 | x = self(x)
18 |
19 | probs = F.softmax(x, dim=-1)
20 | if deterministic is False:
21 | action = probs.multinomial(1)
22 | else:
23 | action = probs.max(1)[1]
24 | return action
25 |
26 | def logprobs_and_entropy(self, x):
27 | x = self(x)
28 |
29 | log_probs = F.log_softmax(x, dim=-1)
30 | probs = F.softmax(x, dim=-1)
31 |
32 | dist_entropy = -(log_probs * probs).sum(-1).mean()
33 | return probs, dist_entropy
34 |
35 |
36 | class Actor(nn.Module):
37 | def __init__(self, num_inputs, action_space, num_outputs):
38 | super(Actor, self).__init__()
39 | self.action_dim = action_space.shape[0]
40 | self.num_outputs = num_outputs
41 |
42 | self.common = nn.Sequential(
43 | nn.Linear(num_inputs, 400),
44 | nn.ReLU(),
45 | nn.Linear(400, 300),
46 | nn.ReLU()
47 | )
48 |
49 | self.mu = nn.Linear(300, self.action_dim * self.num_outputs)
50 | self.dist = Categorical(300, self.num_outputs)
51 |
52 | def forward(self, x):
53 | common = self.common(x)
54 | mu = F.tanh(self.mu(common)).view(x.shape[0], self.num_outputs, self.action_dim)
55 | action = self.dist.sample(common)
56 | probs, dist_entropy = self.dist.logprobs_and_entropy(x)
57 | return mu, action, probs, dist_entropy
58 |
59 |
60 | def cosine_basis_functions(x, n_basis_functions=64):
61 | x = x.view(-1, 1)
62 | i_pi = np.tile(np.arange(1, n_basis_functions + 1, dtype=np.float32), (x.shape[0], 1)) * np.pi
63 | i_pi = torch.Tensor(i_pi)
64 | if x.is_cuda:
65 | i_pi = i_pi.cuda()
66 | embedding = (x * i_pi).cos()
67 | return embedding
68 |
69 |
70 | class CosineBasisLinear(nn.Module):
71 | def __init__(self, n_basis_functions, out_size):
72 | super(CosineBasisLinear, self).__init__()
73 | self.linear = nn.Linear(n_basis_functions, out_size)
74 | self.n_basis_functions = n_basis_functions
75 | self.out_size = out_size
76 |
77 | def forward(self, x):
78 | batch_size = x.shape[0]
79 | h = cosine_basis_functions(x, self.n_basis_functions)
80 | out = self.linear(h)
81 | out = out.view(batch_size, -1, self.out_size)
82 | return out
83 |
84 |
85 | class AutoRegressiveStochasticActor(nn.Module):
86 | def __init__(self, num_inputs, action_dim, n_basis_functions):
87 | super(AutoRegressiveStochasticActor, self).__init__()
88 | self.action_dim = action_dim
89 | self.state_embedding = nn.Linear(num_inputs, 400)
90 | self.noise_embedding = CosineBasisLinear(n_basis_functions, 400)
91 | self.action_embedding = CosineBasisLinear(n_basis_functions, 400)
92 |
93 | self.rnn = nn.GRU(800, 400, batch_first=True)
94 | self.l1 = nn.Linear(400, 400)
95 | self.l2 = nn.Linear(400, 1)
96 |
97 | def forward(self, state, taus, actions=None):
98 | if actions is not None:
99 | return self.supervised_forward(state, taus, actions)
100 | batch_size = state.shape[0]
101 | # batch x 1 x 400
102 | state_embedding = F.leaky_relu(self.state_embedding(state)).unsqueeze(1)
103 | # batch x action dim x 400
104 | noise_embedding = self.noise_embedding(taus)
105 |
106 | action_list = []
107 |
108 | action = torch.zeros(batch_size, 1)
109 | if state.is_cuda:
110 | action = action.cuda()
111 | hidden_state = None
112 |
113 | for idx in range(self.action_dim):
114 | # batch x 1 x 400
115 | action_embedding = F.leaky_relu(self.action_embedding(action.view(batch_size, 1, 1)))
116 | rnn_input = torch.cat([state_embedding, action_embedding], dim=2)
117 | gru_out, hidden_state = self.rnn(rnn_input, hidden_state)
118 |
119 | # batch x 400
120 | hadamard_product = gru_out.squeeze(1) * noise_embedding[:, idx, :]
121 | action = torch.tanh(self.l2(F.leaky_relu(self.l1(hadamard_product))))
122 | action_list.append(action)
123 |
124 | actions = torch.stack(action_list, dim=1).squeeze(-1)
125 | return actions
126 |
127 | def supervised_forward(self, state, taus, actions):
128 | # batch x action dim x 400
129 | state_embedding = F.leaky_relu(self.state_embedding(state)).unsqueeze(1).expand(-1, self.action_dim, -1)
130 | # batch x action dim x 400
131 | shifted_actions = torch.zeros_like(actions)
132 | shifted_actions[:, 1:] = actions[:, :-1]
133 | provided_action_embedding = F.leaky_relu(self.action_embedding(shifted_actions))
134 |
135 | rnn_input = torch.cat([state_embedding, provided_action_embedding], dim=2)
136 | gru_out, _ = self.rnn(rnn_input)
137 |
138 | # batch x action dim x 400
139 | noise_embedding = self.noise_embedding(taus)
140 | # batch x action dim x 400
141 | hadamard_product = gru_out * noise_embedding
142 | actions = torch.tanh(self.l2(F.leaky_relu(self.l1(hadamard_product))))
143 | return actions.squeeze(-1)
144 |
145 |
146 | class StochasticActor(nn.Module):
147 | def __init__(self, num_inputs, action_dim, n_basis_functions):
148 | super(StochasticActor, self).__init__()
149 |
150 | hidden_size = 400
151 |
152 | self.hidden_size = hidden_size
153 | self.action_dim = action_dim
154 | self.l1 = nn.Linear(num_inputs, self.hidden_size)
155 | self.phi = CosineBasisLinear(n_basis_functions, self.hidden_size)
156 | self.l2 = nn.Linear(self.hidden_size, 200)
157 | self.l3 = nn.Linear(200, self.action_dim)
158 |
159 | def forward(self, state, tau, actions):
160 | # batch x ~400
161 | state_embedding = F.leaky_relu(self.l1(state))
162 | # batch x ~400
163 | noise_embedding = F.leaky_relu(self.phi(tau)).view(-1, self.hidden_size)
164 |
165 | hadamard_product = state_embedding * noise_embedding
166 |
167 | l2 = F.leaky_relu(self.l2(hadamard_product))
168 |
169 | actions = torch.tanh(self.l3(l2))
170 |
171 | return actions
172 |
173 |
174 | class Critic(nn.Module):
175 | def __init__(self, num_inputs, num_networks=1):
176 | super(Critic, self).__init__()
177 | self.num_networks = num_networks
178 | self.q1 = nn.Sequential(
179 | nn.Linear(num_inputs, 400),
180 | nn.LeakyReLU(),
181 | nn.Linear(400, 300),
182 | nn.LeakyReLU(),
183 | nn.Linear(300, 1)
184 | )
185 |
186 | if self.num_networks == 2:
187 | self.q2 = nn.Sequential(
188 | nn.Linear(num_inputs, 400),
189 | nn.LeakyReLU(),
190 | nn.Linear(400, 300),
191 | nn.LeakyReLU(),
192 | nn.Linear(300, 1)
193 | )
194 | elif self.num_networks > 2 or self.num_networks < 1:
195 | raise NotImplementedError
196 |
197 | def forward(self, x):
198 | if self.num_networks == 1:
199 | return self.q1(x)
200 | return self.q1(x), self.q2(x)
201 |
--------------------------------------------------------------------------------
/noises/ounoise.py:
--------------------------------------------------------------------------------
1 | import numpy as np
2 |
3 |
4 | # Taken from OpenAI baselines - baselines/ddpg/noise.py
5 |
6 | class ActionNoise(object):
7 | def reset(self):
8 | pass
9 |
10 |
11 | class NormalActionNoise(ActionNoise):
12 | def __init__(self, mu, sigma):
13 | self.mu = mu
14 | self.sigma = sigma
15 |
16 | def __call__(self):
17 | return np.random.normal(self.mu, self.sigma)
18 |
19 | def reset(self):
20 | pass
21 |
22 | def __repr__(self):
23 | return 'NormalActionNoise(mu={}, sigma={})'.format(self.mu, self.sigma)
24 |
25 |
26 | class OrnsteinUhlenbeckActionNoise(ActionNoise):
27 | def __init__(self, mu, sigma, theta=.15, dt=1e-2, x0=None):
28 | self.theta = theta
29 | self.mu = mu
30 | self.sigma = sigma
31 | self.dt = dt
32 | self.x0 = x0
33 | self.reset()
34 |
35 | def __call__(self):
36 | x = self.x_prev + self.theta * (self.mu - self.x_prev) * self.dt + self.sigma * np.sqrt(self.dt) * np.random.normal(size=self.mu.shape)
37 | self.x_prev = x
38 | return x
39 |
40 | def reset(self):
41 | self.x_prev = self.x0 if self.x0 is not None else np.zeros_like(self.mu)
42 |
43 | def __repr__(self):
44 | return 'OrnsteinUhlenbeckActionNoise(mu={}, sigma={})'.format(self.mu, self.sigma)
45 |
--------------------------------------------------------------------------------
/noises/param_noise.py:
--------------------------------------------------------------------------------
1 | import numpy as np
2 | from math import sqrt
3 |
4 | """
5 | From OpenAI Baselines:
6 | https://github.com/openai/baselines/blob/master/baselines/ddpg/noise.py
7 | """
8 |
9 |
10 | class AdaptiveParamNoiseSpec(object):
11 | def __init__(self, initial_stddev=0.1, desired_action_stddev=0.2, adaptation_coefficient=1.01):
12 | """
13 | Note that initial_stddev and current_stddev refer to std of parameter noise,
14 | but desired_action_stddev refers to (as name notes) desired std in action space
15 | """
16 | self.initial_stddev = initial_stddev
17 | self.desired_action_stddev = desired_action_stddev
18 | self.adaptation_coefficient = adaptation_coefficient
19 |
20 | self.current_stddev = initial_stddev
21 |
22 | def adapt(self, distance):
23 | if distance > self.desired_action_stddev:
24 | # Decrease stddev.
25 | self.current_stddev /= self.adaptation_coefficient
26 | else:
27 | # Increase stddev.
28 | self.current_stddev *= self.adaptation_coefficient
29 |
30 | def get_stats(self):
31 | stats = {
32 | 'param_noise_stddev': self.current_stddev,
33 | }
34 | return stats
35 |
36 | def __repr__(self):
37 | fmt = 'AdaptiveParamNoiseSpec(initial_stddev={}, desired_action_stddev={}, adaptation_coefficient={})'
38 | return fmt.format(self.initial_stddev, self.desired_action_stddev, self.adaptation_coefficient)
39 |
40 |
41 | def ddpg_distance_metric(actions1, actions2):
42 | """
43 | Compute "distance" between actions taken by two policies at the same states
44 | Expects numpy arrays
45 | """
46 | diff = actions1-actions2
47 | mean_diff = np.mean(np.square(diff), axis=0)
48 | dist = sqrt(np.mean(mean_diff))
49 | return dist
50 |
--------------------------------------------------------------------------------
/normalized_actions.py:
--------------------------------------------------------------------------------
1 | import gym
2 | import torch
3 |
4 |
5 | class NormalizedActions(gym.ActionWrapper):
6 | def action(self, action):
7 | action = (action + 1) / 2 # [-1, 1] => [0, 1]
8 | action *= (self.action_space.high - self.action_space.low)
9 | action += self.action_space.low
10 | return action
11 |
12 | def _action(self, action):
13 | action = (action + 1) / 2 # [-1, 1] => [0, 1]
14 | action *= (self.action_space.high - self.action_space.low)
15 | action += self.action_space.low
16 | return action
17 |
18 | def _reverse_action(self, action):
19 | action -= self.action_space.low
20 | action /= (self.action_space.high - self.action_space.low)
21 | action = action * 2 - 1
22 | return action
23 |
24 |
25 | def normalize(x, stats):
26 | if stats is None:
27 | return x
28 | return (x - stats.mean) / (stats.var + 1e-8).sqrt()
29 |
30 |
31 | class RunningMeanStd(object):
32 | # https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
33 | def __init__(self, epsilon=1e-4, shape=(), device=torch.device('cpu')):
34 | self.mean = torch.zeros(shape).to(device)
35 | self.var = torch.ones(shape).to(device)
36 | self.count = epsilon
37 |
38 | def update(self, x):
39 | batch_mean = torch.mean(x, dim=0)
40 | batch_var = torch.var(x, dim=0)
41 | batch_count = x.shape[0]
42 | self.update_from_moments(batch_mean, batch_var, batch_count)
43 |
44 | def update_from_moments(self, batch_mean, batch_var, batch_count):
45 | self.mean, self.var, self.count = update_mean_var_count_from_moments(
46 | self.mean, self.var, self.count, batch_mean, batch_var, batch_count)
47 |
48 |
49 | def update_mean_var_count_from_moments(mean, var, count, batch_mean, batch_var, batch_count):
50 | delta = batch_mean - mean
51 | tot_count = count + batch_count
52 |
53 | new_mean = mean + delta * batch_count / tot_count
54 | m_a = var * count
55 | m_b = batch_var * batch_count
56 | M2 = m_a + m_b + delta.sqrt() * count * batch_count / tot_count
57 | new_var = M2 / tot_count
58 | new_count = tot_count
59 |
60 | return new_mean, new_var, new_count
61 |
--------------------------------------------------------------------------------
/policies/generative.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import math
3 | from torch.optim import Adam, SGD
4 | from torch.autograd import Variable
5 | from torch.distributions import Uniform
6 | import torch.nn.functional as F
7 | from network import Critic, StochasticActor, AutoRegressiveStochasticActor
8 | from policies.policy import Policy, hard_update, soft_update
9 |
10 |
11 | def compute_eltwise_huber_quantile_loss(actions, target_actions, taus, weighting):
12 | """Compute elementwise Huber losses for quantile regression.
13 | This is based on Algorithm 1 of https://arxiv.org/abs/1806.06923.
14 | This function assumes that, both of the two kinds of quantile thresholds,
15 | taus (used to compute y) and taus_prime (used to compute t) are iid samples
16 | from U([0,1]).
17 | Args:
18 | actions (Variable): Quantile prediction from taus as a
19 | (batch_size, N, K)-shaped array.
20 | target_actions (Variable): Quantile targets from taus as a
21 | (batch_size, N, K)-shaped array.
22 | taus (ndarray): Quantile thresholds used to compute y as a
23 | (batch_size, N, 1)-shaped array.
24 | Returns:
25 | Variable: Loss
26 | """
27 | I_delta = ((actions - target_actions) > 0).float()
28 | eltwise_huber_loss = F.smooth_l1_loss(actions, target_actions, reduce=False)
29 | eltwise_loss = abs(taus - I_delta) * eltwise_huber_loss * weighting
30 | return eltwise_loss.mean()
31 |
32 |
33 | class Generative(Policy):
34 | def __init__(self, gamma, tau, num_inputs, action_space, replay_size, num_basis_functions=64, actor_samples=1,
35 | q_normalization=0.01, target_policy='linear', target_policy_q='min', autoregressive=True, temp=1.0):
36 |
37 | super(Generative, self).__init__(gamma=gamma, tau=tau, num_inputs=num_inputs, action_space=action_space,
38 | replay_size=replay_size)
39 |
40 | self.actor_samples = actor_samples
41 | self.topk = math.ceil(self.actor_samples * 0.7)
42 | self.num_basis_functions = num_basis_functions
43 | self.action_dim = self.action_space.shape[0]
44 | self.q_normalization = q_normalization
45 | self.target_policy = target_policy
46 | self.autoregressive = autoregressive
47 | self.temp = temp
48 |
49 | if target_policy_q == 'min':
50 | self.target_policy_q = lambda x, y: torch.min(x, y)
51 | elif target_policy_q == 'max':
52 | self.target_policy_q = lambda x, y: torch.max(x, y)
53 | elif target_policy_q == 'mean':
54 | self.target_policy_q = lambda x, y: (x + y / 2)
55 | else:
56 | self.target_policy_q = lambda x, y: x
57 |
58 | self.tau_sampler = Uniform(self.Tensor([0.0]), self.Tensor([1.0]))
59 |
60 | '''
61 | Define networks and optimizers
62 | '''
63 |
64 | if self.autoregressive:
65 | self.actor = AutoRegressiveStochasticActor(self.num_inputs, self.action_dim, self.num_basis_functions).to(self.device)
66 | self.actor_target = AutoRegressiveStochasticActor(self.num_inputs, self.action_dim, self.num_basis_functions).to(self.device)
67 | else:
68 | self.actor = StochasticActor(self.num_inputs, self.action_dim, self.num_basis_functions).to(self.device)
69 | self.actor_target = StochasticActor(self.num_inputs, self.action_dim, self.num_basis_functions).to(self.device)
70 | self.actor_optim = Adam(self.actor.parameters(), lr=1e-4)
71 |
72 | self.critic = Critic(self.num_inputs + self.action_dim, num_networks=2).to(self.device)
73 | self.critic_target = Critic(self.num_inputs + self.action_dim, num_networks=2).to(self.device)
74 | self.critic_optim = Adam(self.critic.parameters(), lr=1e-3)
75 |
76 | self.value = Critic(self.num_inputs).to(self.device)
77 | self.value_target = Critic(self.num_inputs).to(self.device)
78 | self.value_optim = Adam(self.value.parameters(), lr=1e-3)
79 |
80 | '''
81 | For multi-gpu setups we enable data parallelism, due to large sample sizes
82 | '''
83 | if torch.cuda.device_count() > 1:
84 | self.actor = torch.nn.DataParallel(self.actor)
85 | self.actor_target = torch.nn.DataParallel(self.actor_target)
86 |
87 | self.critic = torch.nn.DataParallel(self.critic)
88 | self.critic_target = torch.nn.DataParallel(self.critic_target)
89 |
90 | self.value = torch.nn.DataParallel(self.value)
91 | self.value_target = torch.nn.DataParallel(self.value_target)
92 |
93 | '''
94 | Initialize target network with the same parameters as the main network
95 | '''
96 | hard_update(self.actor_target, self.actor)
97 | hard_update(self.critic_target, self.critic)
98 | hard_update(self.value_target, self.value)
99 |
100 | def eval(self):
101 | self.actor.eval()
102 | self.critic.eval()
103 | self.value.eval()
104 |
105 | def train(self):
106 | self.actor.train()
107 | self.critic.train()
108 | self.value.train()
109 |
110 | def policy(self, actor, state, actions=None):
111 | batch_size = state.shape[0]
112 | '''
113 | We sample a quantile for each dimension of the action.
114 | The action is modeled as an auto-regressive distribution, e.g.,
115 | P(X) = P(x_0) * P(x_1 | x_0) * ... * P(x_n | x_{n-1}, ..., x_0)
116 | '''
117 | if self.autoregressive:
118 | taus = self.tau_sampler.rsample((batch_size, self.action_dim)).view(batch_size, self.action_dim, 1)
119 | else:
120 | taus = self.tau_sampler.rsample((batch_size, 1)).view(batch_size, 1, 1)
121 | return actor(state, taus, actions), None, taus
122 |
123 | def update_critic(self, state_batch, action_batch, reward_batch, mask_batch, next_state_batch):
124 | batch_size = state_batch.shape[0]
125 |
126 | '''
127 | Update value network
128 | '''
129 | with torch.no_grad():
130 | # the value is calculated based on multiple samples from the policy and evaluated using the Q networks
131 | tiled_next_state_batch = self._tile(next_state_batch, 0, self.actor_samples)
132 | tiled_next_action_batch = self.policy(self.actor_target, tiled_next_state_batch)[0].view(batch_size * self.actor_samples, -1)
133 |
134 | next_q1, next_q2 = self.critic_target(torch.cat((tiled_next_state_batch, tiled_next_action_batch), 1))
135 |
136 | # to avoid over-estimation, we use the minimal value calculated between both Q networks
137 | next_v = self.target_policy_q(
138 | (torch.topk(next_q1.view(batch_size, self.actor_samples), self.topk)[0]).mean(-1).unsqueeze(-1),
139 | (torch.topk(next_q2.view(batch_size, self.actor_samples), self.topk)[0]).mean(-1).unsqueeze(-1)
140 | )
141 | v = self.value(state_batch)
142 | value_loss = F.mse_loss(v, next_v)
143 |
144 | with torch.no_grad():
145 | next_v = self.value_target(next_state_batch)
146 | target_q = reward_batch + self.gamma * mask_batch * next_v
147 |
148 | self.value_optim.zero_grad()
149 | value_loss.backward()
150 | torch.nn.utils.clip_grad_norm_(self.value.parameters(), 5.0)
151 | self.value_optim.step()
152 |
153 | '''
154 | Update Q networks
155 | '''
156 |
157 | # Add regularization for the Q function. Similar actions should result in similar Q values.
158 | # noise = (torch.randn_like(action_batch) * self.q_normalization).clamp(-0.5, 0.5)
159 | # action_batch = (action_batch + noise).clamp(-1, 1)
160 |
161 | noise = (self.tau_sampler.rsample((batch_size, self.action_dim)).view(batch_size, self.action_dim) * 2 - 1) * self.q_normalization
162 | action_batch = (action_batch + noise).clamp(-1, 1)
163 |
164 | q1, q2 = self.critic(torch.cat((state_batch, action_batch), 1))
165 | q1_loss = F.mse_loss(q1, target_q)
166 | q2_loss = F.mse_loss(q2, target_q)
167 | critic_loss = q1_loss + q2_loss
168 |
169 | self.critic_optim.zero_grad()
170 | critic_loss.backward()
171 | torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 5.0)
172 | self.critic_optim.step()
173 |
174 | return critic_loss.item() + value_loss.item()
175 |
176 | def update_actor(self, state_batch, action_batch):
177 | batch_size = state_batch.shape[0]
178 | tiled_state_batch = self._tile(state_batch, 0, self.actor_samples)
179 |
180 | with torch.no_grad():
181 | # Calculate the value of each state
182 | values = self.value_target(state_batch)
183 | values = torch.cat([values, values], dim=0)
184 |
185 | '''
186 | Sample multiple actions both from the target policy and from a uniform distribution over the action
187 | space. These samples are used to compute the target distribution, which is defined as all the actions
188 | where Q(state, action) > V(state).
189 | '''
190 | target_actions = self.policy(self.actor_target, tiled_state_batch)[0]
191 | target_actions += torch.randn_like(target_actions) * 0.01
192 | target_actions = target_actions.clamp(-1, 1)
193 |
194 | target_q1, target_q2 = self.critic_target(torch.cat((tiled_state_batch, target_actions), 1))
195 | target_action_values = self.target_policy_q(
196 | target_q1.view(batch_size, self.actor_samples, -1),
197 | target_q2.view(batch_size, self.actor_samples, -1)
198 | )
199 |
200 | random_actions = torch.rand_like(target_actions) * 2 - 1
201 | random_q1, random_q2 = self.critic_target(torch.cat((tiled_state_batch, random_actions), 1))
202 | target_random_values = self.target_policy_q(
203 | random_q1.view(batch_size, self.actor_samples, -1),
204 | random_q2.view(batch_size, self.actor_samples, -1)
205 | )
206 |
207 | target_actions = target_actions.view(batch_size, self.actor_samples, -1)
208 | random_actions = random_actions.view(batch_size, self.actor_samples, -1)
209 |
210 | target_actions = torch.cat([target_actions, random_actions], dim=0)
211 | target_action_values = torch.cat([target_action_values, target_random_values], dim=0)
212 |
213 | # (batch_size, 1) -> (batch_size, N, 1)
214 | values = values.unsqueeze(-1).expand(-1, self.actor_samples, -1)
215 | improvement = (target_action_values > values).view(-1, 1) # Choose everything over value
216 |
217 | weighting_improvement = improvement.view(batch_size * 2, self.actor_samples)
218 | state_improvement = improvement.expand(-1, tiled_state_batch.shape[1])
219 | action_improvement = improvement.expand(-1, self.action_dim)
220 |
221 | tiled_state_batch = torch.cat([tiled_state_batch, tiled_state_batch], dim=0)
222 | improving_state_batch = tiled_state_batch[state_improvement].view(-1, tiled_state_batch.shape[1])
223 | improving_action_batch = target_actions.view(-1, self.action_dim)[action_improvement].view(-1, self.action_dim)
224 |
225 | if self.target_policy == 'linear':
226 | weighting = (target_action_values[weighting_improvement] - values[weighting_improvement])
227 | weighting = weighting / weighting.sum(-1, keepdim=True)
228 | elif self.target_policy == 'exponential':
229 | weighting = torch.exp(target_action_values[weighting_improvement] - values[weighting_improvement])
230 | weighting = torch.clamp(weighting, max=20)
231 | elif self.target_policy == 'boltzman':
232 | weighting = (target_action_values[weighting_improvement] - values[weighting_improvement])
233 | weighting = F.softmax((1./self.temp) * weighting, dim=1)
234 | elif self.target_policy == 'uniform':
235 | weighting = torch.ones_like(target_action_values[weighting_improvement])
236 | else: # argmax
237 | raise NotImplementedError
238 |
239 | if improving_state_batch.shape[0] > 0:
240 | # Sample multiple actions for each state as an estimation of the current policy
241 | actions, _, taus = self.policy(self.actor, improving_state_batch, improving_action_batch)
242 | policy_loss = compute_eltwise_huber_quantile_loss(actions, improving_action_batch, taus.squeeze(-1), weighting)
243 |
244 | self.actor_optim.zero_grad()
245 | policy_loss.backward()
246 | torch.nn.utils.clip_grad_norm_(self.actor.parameters(), 1)
247 | self.actor_optim.step()
248 |
249 | return policy_loss.item()
250 | else:
251 | return 0
252 |
253 | def soft_update(self):
254 | soft_update(self.actor_target, self.actor, self.tau)
255 | soft_update(self.critic_target, self.critic, self.tau)
256 | soft_update(self.value_target, self.value, self.tau)
257 |
--------------------------------------------------------------------------------
/policies/policy.py:
--------------------------------------------------------------------------------
1 | import torch
2 | from torch.autograd import Variable
3 | import os
4 | import numpy as np
5 | from replay_memory import ReplayMemory, Transition
6 |
7 |
8 | def soft_update(target, source, tau):
9 | for target_param, param in zip(target.parameters(), source.parameters()):
10 | target_param.data.copy_(target_param.data * (1.0 - tau) + param.data * tau)
11 |
12 |
13 | def hard_update(target, source):
14 | for target_param, param in zip(target.parameters(), source.parameters()):
15 | target_param.data.copy_(param.data)
16 |
17 |
18 | def get_free_gpu():
19 | os.system('nvidia-smi -q -d Memory |grep -A4 GPU|grep Free >tmp')
20 | memory_available = [int(x.split()[2]) for x in open('tmp', 'r').readlines()]
21 | return np.argmax(memory_available)
22 |
23 |
24 | class Policy:
25 | def __init__(self, gamma, tau, num_inputs, action_space, replay_size):
26 | if torch.cuda.is_available():
27 | self.device = torch.device('cuda')
28 | torch.backends.cudnn.enabled = False
29 | self.Tensor = torch.cuda.FloatTensor
30 | else:
31 | self.device = torch.device('cpu')
32 | self.Tensor = torch.FloatTensor
33 |
34 | self.num_inputs = num_inputs
35 | self.action_space = action_space
36 |
37 | self.gamma = gamma
38 | self.tau = tau
39 |
40 | self.memory = ReplayMemory(replay_size)
41 | self.actor = None
42 |
43 | def eval(self):
44 | raise NotImplementedError
45 |
46 | def train(self):
47 | raise NotImplementedError
48 |
49 | def select_action(self, state, action_noise=None):
50 | state = Variable(state).to(self.device)
51 |
52 | action = self.policy(self.actor, state)[0]
53 |
54 | action = action.data
55 | if action_noise is not None:
56 | action += self.Tensor(action_noise()).to(self.device)
57 |
58 | action = action.clamp(-1, 1)
59 |
60 | return action
61 |
62 | def policy(self, actor, state):
63 | raise NotImplementedError
64 |
65 | def store_transition(self, state, action, mask, next_state, reward):
66 | B = state.shape[0]
67 | for b in range(B):
68 | self.memory.push(state[b], action[b], mask[b], next_state[b], reward[b])
69 |
70 | def update_critic(self, state_batch, action_batch, reward_batch, mask_batch, next_state_batch):
71 | raise NotImplementedError
72 |
73 | def update_actor(self, state_batch, action_batch):
74 | raise NotImplementedError
75 |
76 | def update_parameters(self, batch_size, number_of_iterations):
77 | policy_losses = []
78 | value_losses = []
79 |
80 | for _ in range(number_of_iterations):
81 | transitions = self.memory.sample(batch_size)
82 | batch = Transition(*zip(*transitions))
83 |
84 | state_batch = Variable(torch.stack(batch.state)).to(self.device)
85 | action_batch = Variable(torch.stack(batch.action)).to(self.device)
86 | reward_batch = Variable(torch.stack(batch.reward)).to(self.device).unsqueeze(1)
87 | mask_batch = Variable(torch.stack(batch.mask)).to(self.device).unsqueeze(1)
88 | next_state_batch = Variable(torch.stack(batch.next_state)).to(self.device)
89 |
90 | value_loss = self.update_critic(state_batch, action_batch, reward_batch, mask_batch, next_state_batch)
91 | value_losses.append(value_loss)
92 |
93 | policy_loss = self.update_actor(state_batch, action_batch)
94 | policy_losses.append(policy_loss)
95 | self.soft_update()
96 |
97 | return np.mean(value_losses), np.mean(policy_losses)
98 |
99 | def soft_update(self):
100 | raise NotImplementedError
101 |
102 | def _tile(self, a, dim, n_tile):
103 | init_dim = a.size(dim)
104 | repeat_idx = [1] * a.dim()
105 | repeat_idx[dim] = n_tile
106 | a = a.repeat(*(repeat_idx))
107 | order_index = torch.LongTensor(np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)])).to(
108 | self.device)
109 | return torch.index_select(a, dim, order_index)
110 |
--------------------------------------------------------------------------------
/replay_memory.py:
--------------------------------------------------------------------------------
1 | import random
2 | from collections import namedtuple
3 |
4 | # Taken from
5 | # https://github.com/pytorch/tutorials/blob/master/Reinforcement%20(Q-)Learning%20with%20PyTorch.ipynb
6 |
7 | Transition = namedtuple(
8 | 'Transition', ('state', 'action', 'mask', 'next_state', 'reward'))
9 |
10 |
11 | class ReplayMemory(object):
12 |
13 | def __init__(self, capacity):
14 | self.capacity = capacity
15 | self.memory = []
16 | self.position = 0
17 |
18 | def push(self, *args):
19 | """Saves a transition."""
20 | if len(self.memory) < self.capacity:
21 | self.memory.append(None)
22 | self.memory[self.position] = Transition(*args)
23 | self.position = (self.position + 1) % self.capacity
24 |
25 | def sample(self, batch_size):
26 | return random.sample(self.memory, batch_size)
27 |
28 | def __len__(self):
29 | return len(self.memory)
30 |
--------------------------------------------------------------------------------
/utils.py:
--------------------------------------------------------------------------------
1 | import os
2 | import torch
3 | import numpy as np
4 |
5 |
6 | def save_model(actor, basedir=None):
7 | if not os.path.exists('models/'):
8 | os.makedirs('models/')
9 |
10 | actor_path = "{}/ddpg_actor".format(basedir)
11 | torch.save(actor.state_dict(), actor_path)
12 |
13 |
14 | def load_model(agent, basedir=None):
15 | actor_path = "{}/ddpg_actor".format(basedir)
16 |
17 | print('Loading model from {}'.format(actor_path))
18 | agent.actor.load_state_dict(torch.load(actor_path))
19 |
20 |
21 | def moving_average(a, n=3):
22 | plot_data = np.zeros_like(a)
23 | for idx in range(len(a)):
24 | length = min(idx, n)
25 | plot_data[idx] = a[idx-length:idx+1].mean()
26 | return plot_data
27 |
28 |
29 | def vis_plot(viz, log_dict):
30 | ma_length = 0
31 | if viz is not None:
32 | for field in log_dict:
33 | if len(log_dict[field]) > 0:
34 | _, values = zip(*log_dict[field])
35 |
36 | plot_data = np.array(log_dict[field])
37 | viz.line(X=plot_data[:, 0], Y=moving_average(plot_data[:, 1], ma_length), win=field,
38 | opts=dict(title=field, legend=[field]))
39 |
--------------------------------------------------------------------------------