Repository: llSourcell/A-Guide-to-DeepMinds-StarCraft-AI-Environment Branch: master Commit: cd8bd7ac637d Files: 14 Total size: 69.9 KB Directory structure: gitextract_etjspkov/ ├── A Guide to DeepMind's StarCraft AI Environment.ipynb ├── LICENSE ├── README.md ├── deepq_mineral_shards.py ├── defeat_zerglings/ │ ├── common.py │ ├── demo_agent.py │ ├── dqfd.py │ ├── run_demo_agent.py │ └── train.py ├── enjoy_mineral_shards.py ├── maps/ │ └── chris_maps.py ├── mineral_shards.pkl ├── tests/ │ └── scripted_test.py └── train_mineral_shards.py ================================================ FILE CONTENTS ================================================ ================================================ FILE: A Guide to DeepMind's StarCraft AI Environment.ipynb ================================================ { "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# A Guide to DeepMind's StarCraft AI Environment\n", "\n", "## Demo -- We're going to setup and install the necessary tools to run a pretrained Deep Q Network model on the CollectMineralShards mini-game of DeepMind's StarCraft II Environment.\n", "\n", "![alt text](https://cdn.technologyreview.com/i/images/20161104patrick-strackblizzcon16070a0766.jpg?sw=2360&cx=0&cy=0&cw=950&ch=633 \"Logo Title Text 1\")\n", "\n", "## History\n", "\n", "Deepmind already beat Atari Games with the Deep Q Learner\n", "\n", "![alt text](https://rubenfiszel.github.io/posts/rl4j/conv.png \"Logo Title Text 1\")\n", "![alt text](https://s3-ap-south-1.amazonaws.com/av-blog-media/wp-content/uploads/2017/01/12042140/11038f3.jpg \"Logo Title Text 1\")\n", "![alt text](https://www.intelnervana.com/wp-content/uploads/sites/53/2017/06/Screen-Shot-2015-12-21-at-11.23.43-AM-1.png \"Logo Title Text 1\")\n", "\n", "Then they beat the \"unbeatable\" game of \"Go\" with AlphaGo\n", "\n", "![alt text](https://image.slidesharecdn.com/masteringthegameofgowithdeepneuralnetworksandtreesearch-160321115146/95/mastering-the-game-of-go-with-deep-neural-networks-and-tree-search-10-638.jpg?cb=1458635243 \"Logo Title Text 1\")\n", "![alt text](https://image.slidesharecdn.com/howdeepmindmasteredthegameofgo-160903224536/95/how-deepmind-mastered-the-game-of-go-20-638.jpg?cb=1472943238 \"Logo Title Text 1\")\n", "\n", "And now they've set their sights on Starcraft. For an AI to play StarCraft well, it'll need\n", "\n", "- An effective use of memory\n", "- an ability to plan over a long time\n", "- The capacity to adapt plans based on new information. \n", "- To execute something as simple as “expand your base to some location”, one must coordinate mouse clicks, camera, and available resources. This makes actions and planning hierarchical, which is a challenging aspect of Reinforcement Learning.\n", "\n", "Blizzard's StarCraft II API is an interface that provides full external control of StarCraft II.\n", "\n", "This API exposes functionality for developing software for:\n", "\n", "- Scripted bots.\n", "- Machine-learning based bots.\n", "- Replay analysis.\n", "- Tool assisted human play.\n", "\n", "DeepMind's PySC2 - StarCraft II Learning Environment exposes it as a Python RL Environment. \n", "\n", "- A Machine Learning API developed by Blizzard that gives researchers and developers hooks into the game. This includes the release of tools for Linux for the first time.\n", "- A dataset of anonymised game replays, which will increase from 65k to more than half a million in the coming weeks. \n", "- An open source version of DeepMind’s toolset, PySC2, to allow researchers to easily use Blizzard’s feature-layer API with their agents.\n", "- A series of simple RL mini-games to allow researchers to test the performance of agents on specific tasks.\n", "- A joint paper that outlines the environment, and reports initial baseline results on the mini-games, supervised learning from replays, and the full 1v1 ladder game against the built-in AI.\n", "\n", "Starcraft II is a real-time strategy game developed by Blizzard entertainment, otherwise known as the makers of World of Warcraft. It's the sequel to Starcraft, a game from 1998 that many regard as one of the greatest PC games ever released. Even now, over a decade on, it's still played regularly by people all over the world; in Korea, it's so popular that there are professional leagues dedicated solely to playing the game.\n", "\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Installation Steps\n", "\n", "\n", "Steps\n", "\n", "1) Install pysc2\n", "\n", "2) Clone pysc2-examples repository\n", "\n", "3) Download mini-games StarCraft II Maps\n", "\n", "4) Install Tensorflow, baselines libraries\n", "\n", "5) Open the project with IntelliJ \n", "\n", "6) Run the training script\n", "\n", "7) Run the pre-trained model\n", "\n", "\n", "\n", "## Step 1 - Install pysc2\n", "\n", "`pip3 install pysc2`\n", "\n", "\n", "## Step 2 - Git Clone psyc2 examples\n", "\n", "`git clone https://github.com/llSourcell/A-Guide-to-DeepMinds-StarCraft-AI-Environment`\n", "\n", "## Step 3 - Download mini-games StarCraft II Maps\n", "\n", "https://github.com/deepmind/pysc2/releases/download/v1.0/mini_games.zip\n", "\n", "save these maps to StarCraft II/Maps \n", "\n", "## Step 4 - Install Tensorflow + OpenAI Baselines\n", "\n", "`pip3 install tensorflow`\n", "`pip3 install baselines`\n", "\n", "## Step 5 - Open the Project with Intellij\n", "\n", "### Start training\n", "\n", "`python3 train_mineral_shards.py`\n", "\n", "### Open project , Python 3 SDK \n", "\n", "## Step 6 Run training script\n", "\n", "Right click the train_mineral_shards.py and select [Run 'train_mineral_shards'] menu.\n", "\n", "This is the brief explanation of console logs.\n", "\n", "- steps : The number of commands that we sent to marines.\n", "- episodes : The number of games that we played.\n", "- mean 100 episode reward : mean rewards of last 100 episodes.\n", "- mean 100 episode min… : mean minerals of last 100 episodes.\n", "- % time spent exploring : The percentage of Exploring (Exploration & Exploit)\n", "\n", "\n", "## Step 7 Run pre-trained model\n", "\n", "- Right click the enjoy_mineral_shards.py and select [Run 'enjoy_mineral_shards'] menu.\n", "\n", "Then we can see the pre-trained agent of CollectMineralShards map.\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": true }, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.6.0" } }, "nbformat": 4, "nbformat_minor": 2 } ================================================ FILE: LICENSE ================================================ Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 1. Definitions. "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document. "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License. "Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity. "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License. "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files. "Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types. "Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below). "Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof. "Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." "Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work. 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form. 3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed. 4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions: (a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License. 5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions. 6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file. 7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License. 8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. 9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. END OF TERMS AND CONDITIONS APPENDIX: How to apply the Apache License to your work. To apply the Apache License to your work, attach the following boilerplate notice, with the fields enclosed by brackets "{}" replaced with your own identifying information. (Don't include the brackets!) The text should be enclosed in the appropriate comment syntax for the file format. We also recommend that a file or class name and description of purpose be included on the same "printed page" as the copyright notice for easier identification within third-party archives. Copyright {yyyy} {name of copyright owner} Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ================================================ FILE: README.md ================================================ # A-Guide-to-DeepMind-s-StarCraft-AI-Environment- This is the code for "A Guide to DeepMind's StarCraft AI Environment" by Siraj Raval on Youtube ## Overview This is the code for [this](https://youtu.be/URWXG5jRB-A) video on on Youtube by Siraj Raval. This code will help you train or run a pretrained AI model in the DeepMind Starcraft II environment. ## Dependencies - pysc2 (Deepmind) [https://github.com/deepmind/pysc2] - baselines (OpenAI) [https://github.com/openai/baselines] - s2client-proto (Blizzard) [https://github.com/Blizzard/s2client-proto] - Tensorflow 1.3 (Google) [https://github.com/tensorflow/tensorflow] ## Usage ## 1. Get PySC2 ### PyPI The easiest way to get PySC2 is to use pip: ```shell $ pip install pysc2 ``` Also, you have to install `baselines` library. ```shell $ pip install baselines ``` ## 2. Install StarCraft II ### Mac / Win You have to purchase StarCraft II and install it. Or even the Starter Edition will work. http://us.battle.net/sc2/en/legacy-of-the-void/ ### Linux Packages Follow Blizzard's [documentation](https://github.com/Blizzard/s2client-proto#downloads) to get the linux version. By default, PySC2 expects the game to live in `~/StarCraftII/`. * [3.16.1](http://blzdistsc2-a.akamaihd.net/Linux/SC2.3.16.1.zip) ## 3. Download Maps Download the [ladder maps](https://github.com/Blizzard/s2client-proto#downloads) and the [mini games](https://github.com/deepmind/pysc2/releases/download/v1.0/mini_games.zip) and extract them to your `StarcraftII/Maps/` directory. ## 4. Train it! ```shell $ python train_mineral_shards.py ``` ## 5. Enjoy it! ```shell $ python enjoy_mineral_shards.py ``` ## Credits The credits for this code go to [chris-chris](https://github.com/chris-chris/pysc2-examples). I've merely created a wrapper to get people started. ================================================ FILE: deepq_mineral_shards.py ================================================ import numpy as np import os import dill import tempfile import tensorflow as tf import zipfile import baselines.common.tf_util as U from baselines import logger from baselines.common.schedules import LinearSchedule from baselines import deepq from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer from pysc2.lib import actions as sc2_actions from pysc2.env import environment from pysc2.lib import features from pysc2.lib import actions import gflags as flags _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index _PLAYER_FRIENDLY = 1 _PLAYER_NEUTRAL = 3 # beacon/minerals _PLAYER_HOSTILE = 4 _NO_OP = actions.FUNCTIONS.no_op.id _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _ATTACK_SCREEN = actions.FUNCTIONS.Attack_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _NOT_QUEUED = [0] _SELECT_ALL = [0] FLAGS = flags.FLAGS class ActWrapper(object): def __init__(self, act): self._act = act #self._act_params = act_params @staticmethod def load(path, act_params, num_cpu=16): with open(path, "rb") as f: model_data = dill.load(f) act = deepq.build_act(**act_params) sess = U.make_session(num_cpu=num_cpu) sess.__enter__() with tempfile.TemporaryDirectory() as td: arc_path = os.path.join(td, "packed.zip") with open(arc_path, "wb") as f: f.write(model_data) zipfile.ZipFile(arc_path, 'r', zipfile.ZIP_DEFLATED).extractall(td) U.load_state(os.path.join(td, "model")) return ActWrapper(act) def __call__(self, *args, **kwargs): return self._act(*args, **kwargs) def save(self, path): """Save model to a pickle located at `path`""" with tempfile.TemporaryDirectory() as td: U.save_state(os.path.join(td, "model")) arc_name = os.path.join(td, "packed.zip") with zipfile.ZipFile(arc_name, 'w') as zipf: for root, dirs, files in os.walk(td): for fname in files: file_path = os.path.join(root, fname) if file_path != arc_name: zipf.write(file_path, os.path.relpath(file_path, td)) with open(arc_name, "rb") as f: model_data = f.read() with open(path, "wb") as f: dill.dump((model_data), f) def load(path, act_params, num_cpu=16): """Load act function that was returned by learn function. Parameters ---------- path: str path to the act function pickle num_cpu: int number of cpus to use for executing the policy Returns ------- act: ActWrapper function that takes a batch of observations and returns actions. """ return ActWrapper.load(path, num_cpu=num_cpu, act_params=act_params) def learn(env, q_func, num_actions=4, lr=5e-4, max_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=1, checkpoint_freq=10000, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, num_cpu=16, param_noise=False, param_noise_threshold=0.05, callback=None): """Train a deepq model. Parameters ------- env: pysc2.env.SC2Env environment to train on q_func: (tf.Variable, int, str, bool) -> tf.Variable the model that takes the following inputs: observation_in: object the output of observation placeholder num_actions: int number of actions scope: str reuse: bool should be passed to outer variable scope and returns a tensor of shape (batch_size, num_actions) with values of every action. lr: float learning rate for adam optimizer max_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to max_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. num_cpu: int number of cpus to use for training callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = U.make_session(num_cpu=num_cpu) sess.__enter__() def make_obs_ph(name): return U.BatchInput((64, 64), name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=make_obs_ph, q_func=q_func, num_actions=num_actions, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10 ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': num_actions, } # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = max_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * max_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] #episode_minerals = [0.0] saved_mean_reward = None path_memory = np.zeros((64,64)) obs = env.reset() # Select all marines first obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_ARMY, [_SELECT_ALL])]) player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] screen = player_relative + path_memory player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero() player = [int(player_x.mean()), int(player_y.mean())] if(player[0]>32): screen = shift(LEFT, player[0]-32, screen) elif(player[0]<32): screen = shift(RIGHT, 32 - player[0], screen) if(player[1]>32): screen = shift(UP, player[1]-32, screen) elif(player[1]<32): screen = shift(DOWN, 32 - player[1], screen) reset = True with tempfile.TemporaryDirectory() as td: model_saved = False model_file = os.path.join(td, "model") for t in range(max_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. if param_noise_threshold >= 0.: update_param_noise_threshold = param_noise_threshold else: # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(num_actions)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True action = act(np.array(screen)[None], update_eps=update_eps, **kwargs)[0] reset = False coord = [player[0], player[1]] rew = 0 path_memory_ = np.array(path_memory, copy=True) if(action == 0): #UP if(player[1] >= 16): coord = [player[0], player[1] - 16] path_memory_[player[1] - 16 : player[1], player[0]] = -1 elif(player[1] > 0): coord = [player[0], 0] path_memory_[0 : player[1], player[0]] = -1 #else: # rew -= 1 elif(action == 1): #DOWN if(player[1] <= 47): coord = [player[0], player[1] + 16] path_memory_[player[1] : player[1] + 16, player[0]] = -1 elif(player[1] > 47): coord = [player[0], 63] path_memory_[player[1] : 63, player[0]] = -1 #else: # rew -= 1 elif(action == 2): #LEFT if(player[0] >= 16): coord = [player[0] - 16, player[1]] path_memory_[player[1], player[0] - 16 : player[0]] = -1 elif(player[0] < 16): coord = [0, player[1]] path_memory_[player[1], 0 : player[0]] = -1 #else: # rew -= 1 elif(action == 3): #RIGHT if(player[0] <= 47): coord = [player[0] + 16, player[1]] path_memory_[player[1], player[0] : player[0] + 16] = -1 elif(player[0] > 47): coord = [63, player[1]] path_memory_[player[1], player[0] : 63] = -1 #else: # rew -= 1 #else: #Cannot move, give minus reward # rew -= 1 #if(path_memory[coord[1],coord[0]] != 0): # rew -= 0.5 path_memory = np.array(path_memory_) #print("action : %s Coord : %s" % (action, coord)) if _MOVE_SCREEN not in obs[0].observation["available_actions"]: obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_ARMY, [_SELECT_ALL])]) new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] # else: # new_action = [sc2_actions.FunctionCall(_NO_OP, [])] obs = env.step(actions=new_action) player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] new_screen = player_relative + path_memory player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero() player = [int(player_x.mean()), int(player_y.mean())] if(player[0]>32): new_screen = shift(LEFT, player[0]-32, new_screen) elif(player[0]<32): new_screen = shift(RIGHT, 32 - player[0], new_screen) if(player[1]>32): new_screen = shift(UP, player[1]-32, new_screen) elif(player[1]<32): new_screen = shift(DOWN, 32 - player[1], new_screen) rew = obs[0].reward done = obs[0].step_type == environment.StepType.LAST # Store transition in the replay buffer. replay_buffer.add(screen, action, rew, new_screen, float(done)) screen = new_screen episode_rewards[-1] += rew #episode_minerals[-1] += obs[0].reward if done: obs = env.reset() player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] screen = player_relative + path_memory player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero() player = [int(player_x.mean()), int(player_y.mean())] if(player[0]>32): screen = shift(LEFT, player[0]-32, screen) elif(player[0]<32): screen = shift(RIGHT, 32 - player[0], screen) if(player[1]>32): screen = shift(UP, player[1]-32, screen) elif(player[1]<32): screen = shift(DOWN, 32 - player[1], screen) # Select all marines first env.step(actions=[sc2_actions.FunctionCall(_SELECT_ARMY, [_SELECT_ALL])]) episode_rewards.append(0.0) #episode_minerals.append(0.0) path_memory = np.zeros((64,64)) reset = True if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) #mean_100ep_mineral = round(np.mean(episode_minerals[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len(episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) #logger.record_tabular("mean 100 episode mineral", mean_100ep_mineral) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) U.save_state(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) U.load_state(model_file) return ActWrapper(act) def intToCoordinate(num, size=64): if size!=64: num = num * size * size // 4096 y = num // size x = num - size * y return [x, y] UP, DOWN, LEFT, RIGHT = 'up', 'down', 'left', 'right' def shift(direction, number, matrix): ''' shift given 2D matrix in-place the given number of rows or columns in the specified (UP, DOWN, LEFT, RIGHT) direction and return it ''' if direction in (UP): matrix = np.roll(matrix, -number, axis=0) matrix[number:,:] = -2 return matrix elif direction in (DOWN): matrix = np.roll(matrix, number, axis=0) matrix[:number,:] = -2 return matrix elif direction in (LEFT): matrix = np.roll(matrix, -number, axis=1) matrix[:,number:] = -2 return matrix elif direction in (RIGHT): matrix = np.roll(matrix, number, axis=1) matrix[:,:number] = -2 return matrix else: return matrix ================================================ FILE: defeat_zerglings/common.py ================================================ import numpy as np from pysc2.lib import actions as sc2_actions from pysc2.lib import features from pysc2.lib import actions _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index _UNIT_TYPE = features.SCREEN_FEATURES.unit_type.index _SELECTED = features.SCREEN_FEATURES.selected.index _PLAYER_FRIENDLY = 1 _PLAYER_NEUTRAL = 3 # beacon/minerals _PLAYER_HOSTILE = 4 _NO_OP = actions.FUNCTIONS.no_op.id _SELECT_UNIT_ID = 1 _CONTROL_GROUP_SET = 1 _CONTROL_GROUP_RECALL = 0 _SELECT_CONTROL_GROUP = actions.FUNCTIONS.select_control_group.id _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _ATTACK_SCREEN = actions.FUNCTIONS.Attack_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_UNIT = actions.FUNCTIONS.select_unit.id _SELECT_POINT = actions.FUNCTIONS.select_point.id _NOT_QUEUED = [0] _SELECT_ALL = [0] def init(env, player_relative, obs): #print("init") army_count = env._obs.observation.player_common.army_count if(army_count==0): return obs try: obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero() obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_ARMY, [_SELECT_ALL])]) except Exception as e: print(e) for i in range(len(player_x)): if i % 4 != 0: continue xy = [player_x[i], player_y[i]] obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[0], xy])]) group_id = 0 group_list = [] unit_xy_list = [] for i in range(len(player_x)): if i % 4 != 0: continue if group_id > 9: break xy = [player_x[i], player_y[i]] unit_xy_list.append(xy) if(len(unit_xy_list) >= 1): for idx, xy in enumerate(unit_xy_list): if(idx==0): obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[0], xy])]) else: obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[1], xy])]) obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_CONTROL_GROUP, [[_CONTROL_GROUP_SET], [group_id]])]) unit_xy_list = [] group_list.append(group_id) group_id += 1 if(len(unit_xy_list) >= 1): for idx, xy in enumerate(unit_xy_list): if(idx==0): obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[0], xy])]) else: obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[1], xy])]) obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_CONTROL_GROUP, [[_CONTROL_GROUP_SET], [group_id]])]) group_list.append(group_id) group_id += 1 return obs def update_group_list(obs): control_groups = obs[0].observation["control_groups"] group_count = 0 group_list = [] for id, group in enumerate(control_groups): if(group[0]!=0): group_count += 1 group_list.append(id) return group_list def check_group_list(env, obs): error = False control_groups = obs[0].observation["control_groups"] army_count = 0 for id, group in enumerate(control_groups): if(group[0]==48): army_count += group[1] if(group[1] != 1): #print("group error group_id : %s count : %s" % (id, group[1])) error = True return error if(army_count != env._obs.observation.player_common.army_count): error = True # print("army_count %s != %s env._obs.observation.player_common.army_count " # % (army_count, env._obs.observation.player_common.army_count)) return error UP, DOWN, LEFT, RIGHT = 'up', 'down', 'left', 'right' def shift(direction, number, matrix): ''' shift given 2D matrix in-place the given number of rows or columns in the specified (UP, DOWN, LEFT, RIGHT) direction and return it ''' if direction in (UP): matrix = np.roll(matrix, -number, axis=0) matrix[number:,:] = -2 return matrix elif direction in (DOWN): matrix = np.roll(matrix, number, axis=0) matrix[:number,:] = -2 return matrix elif direction in (LEFT): matrix = np.roll(matrix, -number, axis=1) matrix[:,number:] = -2 return matrix elif direction in (RIGHT): matrix = np.roll(matrix, number, axis=1) matrix[:,:number] = -2 return matrix else: return matrix def select_marine(env, obs): player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] screen = player_relative group_list = update_group_list(obs) if(check_group_list(env, obs)): obs = init(env, player_relative, obs) group_list = update_group_list(obs) # if(len(group_list) == 0): # obs = init(env, player_relative, obs) # group_list = update_group_list(obs) player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] friendly_y, friendly_x = (player_relative == _PLAYER_FRIENDLY).nonzero() enemy_y, enemy_x = (player_relative == _PLAYER_HOSTILE).nonzero() player = [] danger_closest, danger_min_dist = None, None for e in zip(enemy_x, enemy_y): for p in zip(friendly_x, friendly_y): dist = np.linalg.norm(np.array(p) - np.array(e)) if not danger_min_dist or dist < danger_min_dist: danger_closest, danger_min_dist = p, dist marine_closest, marine_min_dist = None, None for e in zip(friendly_x, friendly_y): for p in zip(friendly_x, friendly_y): dist = np.linalg.norm(np.array(p) - np.array(e)) if not marine_min_dist or dist < marine_min_dist: if dist >= 2: marine_closest, marine_min_dist = p, dist if(danger_min_dist != None and danger_min_dist <= 5): obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[0], danger_closest])]) selected = obs[0].observation["screen"][_SELECTED] player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero() if(len(player_y)>0): player = [int(player_x.mean()), int(player_y.mean())] elif(marine_closest != None and marine_min_dist <= 3): obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_POINT, [[0], marine_closest])]) selected = obs[0].observation["screen"][_SELECTED] player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero() if(len(player_y)>0): player = [int(player_x.mean()), int(player_y.mean())] else: # If there is no marine in danger, select random while(len(group_list)>0): # units = env._obs.observation.raw_data.units # marine_list = [] # for unit in units: # if(unit.alliance == 1): # marine_list.append(unit) group_id = np.random.choice(group_list) #xy = [int(unit.pos.y - 10), int(unit.pos.x+8)] #print("check xy : %s - %s" % (xy, player_relative[xy[0],xy[1]])) obs = env.step(actions=[sc2_actions.FunctionCall(_SELECT_CONTROL_GROUP, [[_CONTROL_GROUP_RECALL], [group_id]])]) selected = obs[0].observation["screen"][_SELECTED] player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero() if(len(player_y)>0): player = [int(player_x.mean()), int(player_y.mean())] break else: group_list.remove(group_id) if(len(player) == 2): if(player[0]>32): screen = shift(LEFT, player[0]-32, screen) elif(player[0]<32): screen = shift(RIGHT, 32 - player[0], screen) if(player[1]>32): screen = shift(UP, player[1]-32, screen) elif(player[1]<32): screen = shift(DOWN, 32 - player[1], screen) return obs, screen, player def marine_action(env, obs, player, action): player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] enemy_y, enemy_x = (player_relative == _PLAYER_HOSTILE).nonzero() closest, min_dist = None, None if(len(player) == 2): for p in zip(enemy_x, enemy_y): dist = np.linalg.norm(np.array(player) - np.array(p)) if not min_dist or dist < min_dist: closest, min_dist = p, dist player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] friendly_y, friendly_x = (player_relative == _PLAYER_FRIENDLY).nonzero() closest_friend, min_dist_friend = None, None if(len(player) == 2): for p in zip(friendly_x, friendly_y): dist = np.linalg.norm(np.array(player) - np.array(p)) if not min_dist_friend or dist < min_dist_friend: closest_friend, min_dist_friend = p, dist if(closest == None): new_action = [sc2_actions.FunctionCall(_NO_OP, [])] elif(action == 0 and closest_friend != None and min_dist_friend < 3): # Friendly marine is too close => Sparse! mean_friend = [int(friendly_x.mean()), int(friendly_x.mean())] diff = np.array(player) - np.array(closest_friend) norm = np.linalg.norm(diff) if(norm != 0): diff = diff / norm coord = np.array(player) + diff * 4 if(coord[0]<0): coord[0] = 0 elif(coord[0]>63): coord[0] = 63 if(coord[1]<0): coord[1] = 0 elif(coord[1]>63): coord[1] = 63 new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] elif(action <= 1): #Attack # nearest enemy coord = closest new_action = [sc2_actions.FunctionCall(_ATTACK_SCREEN, [_NOT_QUEUED, coord])] #print("action : %s Attack Coord : %s" % (action, coord)) elif(action == 2): # Oppsite direcion from enemy # nearest enemy opposite diff = np.array(player) - np.array(closest) norm = np.linalg.norm(diff) if(norm != 0): diff = diff / norm coord = np.array(player) + diff * 7 if(coord[0]<0): coord[0] = 0 elif(coord[0]>63): coord[0] = 63 if(coord[1]<0): coord[1] = 0 elif(coord[1]>63): coord[1] = 63 new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] elif(action == 4): #UP coord = [player[0], player[1] - 3] new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] elif(action == 5): #DOWN coord = [player[0], player[1] + 3] new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] elif(action == 6): #LEFT coord = [player[0] - 3, player[1]] new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] elif(action == 7): #RIGHT coord = [player[0] + 3, player[1]] new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] #print("action : %s Back Coord : %s" % (action, coord)) return obs, new_action ================================================ FILE: defeat_zerglings/demo_agent.py ================================================ """A random agent for starcraft.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy from pysc2.agents import base_agent from pysc2.lib import actions from pysc2.lib import actions as sc2_actions from pysc2.lib import features from pysc2.lib import actions from defeat_zerglings import common import numpy as np _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index _UNIT_TYPE = features.SCREEN_FEATURES.unit_type.index _SELECTED = features.SCREEN_FEATURES.selected.index _PLAYER_FRIENDLY = 1 _PLAYER_NEUTRAL = 3 # beacon/minerals _PLAYER_HOSTILE = 4 _NO_OP = actions.FUNCTIONS.no_op.id _SELECT_UNIT_ID = 1 _CONTROL_GROUP_SET = 1 _CONTROL_GROUP_RECALL = 0 _SELECT_CONTROL_GROUP = actions.FUNCTIONS.select_control_group.id _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _ATTACK_SCREEN = actions.FUNCTIONS.Attack_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_UNIT = actions.FUNCTIONS.select_unit.id _SELECT_POINT = actions.FUNCTIONS.select_point.id _NOT_QUEUED = [0] _SELECT_ALL = [0] class MarineAgent(base_agent.BaseAgent): """A random agent for starcraft.""" demo_replay = [] def __init__(self, env): self.env = env def step(self, obs): super(MarineAgent, self).step(obs) #1. Select marine! obs, screen, player = common.select_marine(self.env, [obs]) player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] enemy_y, enemy_x = (player_relative == _PLAYER_HOSTILE).nonzero() #2. Run away from nearby enemy closest, min_dist = None, None if(len(player) == 2): for p in zip(enemy_x, enemy_y): dist = np.linalg.norm(np.array(player) - np.array(p)) if not min_dist or dist < min_dist: closest, min_dist = p, dist #3. Sparse! friendly_y, friendly_x = (player_relative == _PLAYER_FRIENDLY).nonzero() closest_friend, min_dist_friend = None, None if(len(player) == 2): for p in zip(friendly_x, friendly_y): dist = np.linalg.norm(np.array(player) - np.array(p)) if not min_dist_friend or dist < min_dist_friend: closest_friend, min_dist_friend = p, dist if(min_dist != None and min_dist <= 7): obs, new_action = common.marine_action(self.env, obs, player, 2) elif(min_dist_friend != None and min_dist_friend <= 3): sparse_or_attack = np.random.randint(0,2) obs, new_action = common.marine_action(self.env, obs, player, sparse_or_attack) else: obs, new_action = common.marine_action(self.env, obs, player, 1) return new_action[0] ================================================ FILE: defeat_zerglings/dqfd.py ================================================ import numpy as np import os import dill import tempfile import tensorflow as tf import zipfile import baselines.common.tf_util as U from baselines import logger from baselines.common.schedules import LinearSchedule from baselines import deepq from baselines.deepq.replay_buffer import ReplayBuffer, PrioritizedReplayBuffer from pysc2.lib import actions as sc2_actions from pysc2.env import environment from pysc2.lib import features from pysc2.lib import actions from defeat_zerglings import common import gflags as flags _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index _UNIT_TYPE = features.SCREEN_FEATURES.unit_type.index _SELECTED = features.SCREEN_FEATURES.selected.index _PLAYER_FRIENDLY = 1 _PLAYER_NEUTRAL = 3 # beacon/minerals _PLAYER_HOSTILE = 4 _NO_OP = actions.FUNCTIONS.no_op.id _SELECT_UNIT_ID = 1 _CONTROL_GROUP_SET = 1 _CONTROL_GROUP_RECALL = 0 _SELECT_CONTROL_GROUP = actions.FUNCTIONS.select_control_group.id _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _ATTACK_SCREEN = actions.FUNCTIONS.Attack_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_UNIT = actions.FUNCTIONS.select_unit.id _SELECT_POINT = actions.FUNCTIONS.select_point.id _NOT_QUEUED = [0] _SELECT_ALL = [0] UP, DOWN, LEFT, RIGHT = 'up', 'down', 'left', 'right' FLAGS = flags.FLAGS class ActWrapper(object): def __init__(self, act): self._act = act #self._act_params = act_params @staticmethod def load(path, act_params, num_cpu=16): with open(path, "rb") as f: model_data = dill.load(f) act = deepq.build_act(**act_params) sess = U.make_session(num_cpu=num_cpu) sess.__enter__() with tempfile.TemporaryDirectory() as td: arc_path = os.path.join(td, "packed.zip") with open(arc_path, "wb") as f: f.write(model_data) zipfile.ZipFile(arc_path, 'r', zipfile.ZIP_DEFLATED).extractall(td) U.load_state(os.path.join(td, "model")) return ActWrapper(act) def __call__(self, *args, **kwargs): return self._act(*args, **kwargs) def save(self, path): """Save model to a pickle located at `path`""" with tempfile.TemporaryDirectory() as td: U.save_state(os.path.join(td, "model")) arc_name = os.path.join(td, "packed.zip") with zipfile.ZipFile(arc_name, 'w') as zipf: for root, dirs, files in os.walk(td): for fname in files: file_path = os.path.join(root, fname) if file_path != arc_name: zipf.write(file_path, os.path.relpath(file_path, td)) with open(arc_name, "rb") as f: model_data = f.read() with open(path, "wb") as f: dill.dump((model_data), f) def load(path, act_params, num_cpu=16): """Load act function that was returned by learn function. Parameters ---------- path: str path to the act function pickle num_cpu: int number of cpus to use for executing the policy Returns ------- act: ActWrapper function that takes a batch of observations and returns actions. """ return ActWrapper.load(path, num_cpu=num_cpu, act_params=act_params) def learn(env, q_func, num_actions=3, lr=5e-4, max_timesteps=100000, buffer_size=50000, exploration_fraction=0.1, exploration_final_eps=0.02, train_freq=1, batch_size=32, print_freq=1, checkpoint_freq=10000, learning_starts=1000, gamma=1.0, target_network_update_freq=500, prioritized_replay=False, prioritized_replay_alpha=0.6, prioritized_replay_beta0=0.4, prioritized_replay_beta_iters=None, prioritized_replay_eps=1e-6, num_cpu=16, param_noise=False, param_noise_threshold=0.05, callback=None, demo_replay=[] ): """Train a deepq model. Parameters ------- env: pysc2.env.SC2Env environment to train on q_func: (tf.Variable, int, str, bool) -> tf.Variable the model that takes the following inputs: observation_in: object the output of observation placeholder num_actions: int number of actions scope: str reuse: bool should be passed to outer variable scope and returns a tensor of shape (batch_size, num_actions) with values of every action. lr: float learning rate for adam optimizer max_timesteps: int number of env steps to optimizer for buffer_size: int size of the replay buffer exploration_fraction: float fraction of entire training period over which the exploration rate is annealed exploration_final_eps: float final value of random action probability train_freq: int update the model every `train_freq` steps. set to None to disable printing batch_size: int size of a batched sampled from replay buffer for training print_freq: int how often to print out training progress set to None to disable printing checkpoint_freq: int how often to save the model. This is so that the best version is restored at the end of the training. If you do not wish to restore the best version at the end of the training set this variable to None. learning_starts: int how many steps of the model to collect transitions for before learning starts gamma: float discount factor target_network_update_freq: int update the target network every `target_network_update_freq` steps. prioritized_replay: True if True prioritized replay buffer will be used. prioritized_replay_alpha: float alpha parameter for prioritized replay buffer prioritized_replay_beta0: float initial value of beta for prioritized replay buffer prioritized_replay_beta_iters: int number of iterations over which beta will be annealed from initial value to 1.0. If set to None equals to max_timesteps. prioritized_replay_eps: float epsilon to add to the TD errors when updating priorities. num_cpu: int number of cpus to use for training callback: (locals, globals) -> None function called at every steps with state of the algorithm. If callback returns true training stops. Returns ------- act: ActWrapper Wrapper over act function. Adds ability to save it and load it. See header of baselines/deepq/categorical.py for details on the act function. """ # Create all the functions necessary to train the model sess = U.make_session(num_cpu=num_cpu) sess.__enter__() def make_obs_ph(name): return U.BatchInput((64, 64), name=name) act, train, update_target, debug = deepq.build_train( make_obs_ph=make_obs_ph, q_func=q_func, num_actions=num_actions, optimizer=tf.train.AdamOptimizer(learning_rate=lr), gamma=gamma, grad_norm_clipping=10 ) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': q_func, 'num_actions': num_actions, } # Create the replay buffer if prioritized_replay: replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha) if prioritized_replay_beta_iters is None: prioritized_replay_beta_iters = max_timesteps beta_schedule = LinearSchedule(prioritized_replay_beta_iters, initial_p=prioritized_replay_beta0, final_p=1.0) else: replay_buffer = ReplayBuffer(buffer_size) beta_schedule = None # Create the schedule for exploration starting from 1. exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * max_timesteps), initial_p=1.0, final_p=exploration_final_eps) # Initialize the parameters and copy them to the target network. U.initialize() update_target() episode_rewards = [0.0] saved_mean_reward = None obs = env.reset() # Select all marines first player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] screen = player_relative obs = common.init(env, player_relative, obs) group_id = 0 reset = True with tempfile.TemporaryDirectory() as td: model_saved = False model_file = os.path.join(td, "model") for t in range(max_timesteps): if callback is not None: if callback(locals(), globals()): break # Take action and update exploration to the newest value kwargs = {} if not param_noise: update_eps = exploration.value(t) update_param_noise_threshold = 0. else: update_eps = 0. if param_noise_threshold >= 0.: update_param_noise_threshold = param_noise_threshold else: # Compute the threshold such that the KL divergence between perturbed and non-perturbed # policy is comparable to eps-greedy exploration with eps = exploration.value(t). # See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017 # for detailed explanation. update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(num_actions)) kwargs['reset'] = reset kwargs['update_param_noise_threshold'] = update_param_noise_threshold kwargs['update_param_noise_scale'] = True # custom process for DefeatZerglingsAndBanelings obs, screen, player = common.select_marine(env, obs) action = act(np.array(screen)[None], update_eps=update_eps, **kwargs)[0] reset = False rew = 0 new_action = None obs, new_action = common.marine_action(env, obs, player, action) army_count = env._obs.observation.player_common.army_count try: if army_count > 0 and _ATTACK_SCREEN in obs[0].observation["available_actions"]: obs = env.step(actions=new_action) else: new_action = [sc2_actions.FunctionCall(_NO_OP, [])] obs = env.step(actions=new_action) except Exception as e: #print(e) 1 # Do nothing player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] new_screen = player_relative rew += obs[0].reward done = obs[0].step_type == environment.StepType.LAST selected = obs[0].observation["screen"][_SELECTED] player_y, player_x = (selected == _PLAYER_FRIENDLY).nonzero() if(len(player_y)>0): player = [int(player_x.mean()), int(player_y.mean())] if(len(player) == 2): if(player[0]>32): new_screen = common.shift(LEFT, player[0]-32, new_screen) elif(player[0]<32): new_screen = common.shift(RIGHT, 32 - player[0], new_screen) if(player[1]>32): new_screen = common.shift(UP, player[1]-32, new_screen) elif(player[1]<32): new_screen = common.shift(DOWN, 32 - player[1], new_screen) # Store transition in the replay buffer. replay_buffer.add(screen, action, rew, new_screen, float(done)) screen = new_screen episode_rewards[-1] += rew if done: print("Episode Reward : %s" % episode_rewards[-1]) obs = env.reset() player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] screen = player_relative group_list = common.init(env, player_relative, obs) # Select all marines first #env.step(actions=[sc2_actions.FunctionCall(_SELECT_UNIT, [_SELECT_ALL])]) episode_rewards.append(0.0) reset = True if t > learning_starts and t % train_freq == 0: # Minimize the error in Bellman's equation on a batch sampled from replay buffer. if prioritized_replay: experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t)) (obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience else: obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size) weights, batch_idxes = np.ones_like(rewards), None td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights) if prioritized_replay: new_priorities = np.abs(td_errors) + prioritized_replay_eps replay_buffer.update_priorities(batch_idxes, new_priorities) if t > learning_starts and t % target_network_update_freq == 0: # Update target network periodically. update_target() mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1) num_episodes = len(episode_rewards) if done and print_freq is not None and len(episode_rewards) % print_freq == 0: logger.record_tabular("steps", t) logger.record_tabular("episodes", num_episodes) logger.record_tabular("mean 100 episode reward", mean_100ep_reward) logger.record_tabular("% time spent exploring", int(100 * exploration.value(t))) logger.dump_tabular() if (checkpoint_freq is not None and t > learning_starts and num_episodes > 100 and t % checkpoint_freq == 0): if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward: if print_freq is not None: logger.log("Saving model due to mean reward increase: {} -> {}".format( saved_mean_reward, mean_100ep_reward)) U.save_state(model_file) model_saved = True saved_mean_reward = mean_100ep_reward if model_saved: if print_freq is not None: logger.log("Restored model with mean reward: {}".format(saved_mean_reward)) U.load_state(model_file) return ActWrapper(act) ================================================ FILE: defeat_zerglings/run_demo_agent.py ================================================ import sys import gflags as flags from baselines import deepq from pysc2.env import sc2_env from pysc2.lib import actions from pysc2.env import run_loop from defeat_zerglings import demo_agent from maps import chris_maps _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_ALL = [0] _NOT_QUEUED = [0] step_mul = 1 steps = 20000 FLAGS = flags.FLAGS def main(): FLAGS(sys.argv) with sc2_env.SC2Env( "DefeatZerglingsAndBanelings", step_mul=step_mul, visualize=True, game_steps_per_episode=steps * step_mul) as env: demo_replay = [] agent = demo_agent.MarineAgent(env=env) agent.env = env run_loop.run_loop([agent], env, steps) if __name__ == '__main__': main() ================================================ FILE: defeat_zerglings/train.py ================================================ import sys import gflags as flags from baselines import deepq from pysc2.env import sc2_env from pysc2.lib import actions from defeat_zerglings import dqfd _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_ALL = [0] _NOT_QUEUED = [0] step_mul = 1 steps = 2000 FLAGS = flags.FLAGS def main(): FLAGS(sys.argv) with sc2_env.SC2Env( "DefeatZerglingsAndBanelings", step_mul=step_mul, visualize=True, game_steps_per_episode=steps * step_mul) as env: model = deepq.models.cnn_to_mlp( convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)], hiddens=[256], dueling=True ) demo_replay = [] act = dqfd.learn( env, q_func=model, num_actions=3, lr=1e-4, max_timesteps=10000000, buffer_size=100000, exploration_fraction=0.5, exploration_final_eps=0.01, train_freq=2, learning_starts=100000, target_network_update_freq=1000, gamma=0.99, prioritized_replay=True, demo_replay=demo_replay ) act.save("defeat_zerglings.pkl") if __name__ == '__main__': main() ================================================ FILE: enjoy_mineral_shards.py ================================================ import sys import baselines.common.tf_util as U import gflags as flags import numpy as np from baselines import deepq from pysc2.env import environment from pysc2.env import sc2_env from pysc2.lib import actions from pysc2.lib import actions as sc2_actions from pysc2.lib import features import deepq_mineral_shards _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index _PLAYER_FRIENDLY = 1 _PLAYER_NEUTRAL = 3 # beacon/minerals _PLAYER_HOSTILE = 4 _NO_OP = actions.FUNCTIONS.no_op.id _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _ATTACK_SCREEN = actions.FUNCTIONS.Attack_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _NOT_QUEUED = [0] _SELECT_ALL = [0] step_mul = 16 steps = 400 FLAGS = flags.FLAGS def main(): FLAGS(sys.argv) with sc2_env.SC2Env( "CollectMineralShards", step_mul=step_mul, visualize=True, game_steps_per_episode=steps * step_mul) as env: model = deepq.models.cnn_to_mlp( convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)], hiddens=[256], dueling=True ) def make_obs_ph(name): return U.BatchInput((64, 64), name=name) act_params = { 'make_obs_ph': make_obs_ph, 'q_func': model, 'num_actions': 4, } act = deepq_mineral_shards.load("mineral_shards.pkl", act_params=act_params) while True: obs = env.reset() episode_rew = 0 done = False step_result = env.step(actions=[sc2_actions.FunctionCall(_SELECT_ARMY, [_SELECT_ALL])]) while not done: player_relative = step_result[0].observation["screen"][_PLAYER_RELATIVE] obs = player_relative player_y, player_x = (player_relative == _PLAYER_FRIENDLY).nonzero() player = [int(player_x.mean()), int(player_y.mean())] if(player[0]>32): obs = shift(LEFT, player[0]-32, obs) elif(player[0]<32): obs = shift(RIGHT, 32 - player[0], obs) if(player[1]>32): obs = shift(UP, player[1]-32, obs) elif(player[1]<32): obs = shift(DOWN, 32 - player[1], obs) action = act(obs[None])[0] coord = [player[0], player[1]] if(action == 0): #UP if(player[1] >= 16): coord = [player[0], player[1] - 16] elif(player[1] > 0): coord = [player[0], 0] elif(action == 1): #DOWN if(player[1] <= 47): coord = [player[0], player[1] + 16] elif(player[1] > 47): coord = [player[0], 63] elif(action == 2): #LEFT if(player[0] >= 16): coord = [player[0] - 16, player[1]] elif(player[0] < 16): coord = [0, player[1]] elif(action == 3): #RIGHT if(player[0] <= 47): coord = [player[0] + 16, player[1]] elif(player[0] > 47): coord = [63, player[1]] new_action = [sc2_actions.FunctionCall(_MOVE_SCREEN, [_NOT_QUEUED, coord])] step_result = env.step(actions=new_action) rew = step_result[0].reward done = step_result[0].step_type == environment.StepType.LAST episode_rew += rew print("Episode reward", episode_rew) UP, DOWN, LEFT, RIGHT = 'up', 'down', 'left', 'right' def shift(direction, number, matrix): ''' shift given 2D matrix in-place the given number of rows or columns in the specified (UP, DOWN, LEFT, RIGHT) direction and return it ''' if direction in (UP): matrix = np.roll(matrix, -number, axis=0) matrix[number:,:] = -2 return matrix elif direction in (DOWN): matrix = np.roll(matrix, number, axis=0) matrix[:number,:] = -2 return matrix elif direction in (LEFT): matrix = np.roll(matrix, -number, axis=1) matrix[:,number:] = -2 return matrix elif direction in (RIGHT): matrix = np.roll(matrix, number, axis=1) matrix[:,:number] = -2 return matrix else: return matrix if __name__ == '__main__': main() ================================================ FILE: maps/chris_maps.py ================================================ """Define the mini game map configs. These are maps made by Deepmind.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function from pysc2.maps import lib class ChrisMaps(lib.Map): directory = "chris_maps" download = "https://github.com/chris-chris/pysc2-examples#get-the-maps" players = 1 score_index = 0 game_steps_per_episode = 0 step_mul = 8 chris_maps = [ "DefeatZealots", # 120s ] for name in chris_maps: globals()[name] = type(name, (ChrisMaps,), dict(filename=name)) ================================================ FILE: tests/scripted_test.py ================================================ #!/usr/bin/python from __future__ import absolute_import from __future__ import division from __future__ import print_function from pysc2.agents import random_agent from pysc2.env import run_loop from pysc2.env import sc2_env from pysc2.tests import utils from pysc2.lib import actions as sc2_actions from pysc2.lib import features from pysc2.lib import basetest import gflags as flags import sys _NO_OP = sc2_actions.FUNCTIONS.no_op.id _PLAYER_RELATIVE = features.SCREEN_FEATURES.player_relative.index FLAGS = flags.FLAGS class TestScripted(utils.TestCase): steps = 2000 step_mul = 1 def test_defeat_zerglings(self): FLAGS(sys.argv) with sc2_env.SC2Env( "DefeatZerglingsAndBanelings", step_mul=self.step_mul, visualize=True, game_steps_per_episode=self.steps * self.step_mul) as env: obs = env.step(actions=[sc2_actions.FunctionCall(_NO_OP, [])]) player_relative = obs[0].observation["screen"][_PLAYER_RELATIVE] # Break Point!! print(player_relative) agent = random_agent.RandomAgent() run_loop.run_loop([agent], env, self.steps) self.assertEqual(agent.steps, self.steps) if __name__ == "__main__": basetest.main() ================================================ FILE: train_mineral_shards.py ================================================ import sys import gflags as flags from baselines import deepq from pysc2.env import sc2_env from pysc2.lib import actions import deepq_mineral_shards _MOVE_SCREEN = actions.FUNCTIONS.Move_screen.id _SELECT_ARMY = actions.FUNCTIONS.select_army.id _SELECT_ALL = [0] _NOT_QUEUED = [0] step_mul = 8 FLAGS = flags.FLAGS def main(): FLAGS(sys.argv) with sc2_env.SC2Env( "CollectMineralShards", step_mul=step_mul, visualize=True) as env: model = deepq.models.cnn_to_mlp( convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)], hiddens=[256], dueling=True ) act = deepq_mineral_shards.learn( env, q_func=model, num_actions=4, lr=1e-5, max_timesteps=2000000, buffer_size=100000, exploration_fraction=0.5, exploration_final_eps=0.01, train_freq=4, learning_starts=100000, target_network_update_freq=1000, gamma=0.99, prioritized_replay=True ) act.save("mineral_shards.pkl") if __name__ == '__main__': main()