Initial commit. First policy extraction on viper works.

.gitignore

@@ -0,0 +1,162 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+log/*

README.md

@@ -0,0 +1,38 @@
+# Viper
+
+Read the accompanying blog post here (tbd).
+
+**V**erifiability via **I**terative **P**olicy **E**xt**R**action (2019) [paper](https://arxiv.org/abs/1805.08328)] 
+
+In this paper the authors distill a Deep Reinforcement Learning such as DeepQN into a decision tree policy which can then be automatically checked for correctness, robustness, and stability. 
+
+This repository implements and tests the viper algorithm on the following environments:
+
+- CartPole
+- Atari Pong
+- ToyPong (tbd)
+
+## Usage
+
+### Training the oracle
+
+Atari Pong:
+
+```
+python main.py train-oracle --env-name PongNoFrameskip-v4 --n-env 8 --total-timesteps 10_000_000
+```
+
+Cart pole:
+
+```
+python main.py train-oracle --env-name CartPole-v1 --n-env 8 --total-timesteps 100_000
+```
+
+
+### Running viper
+
+Once the oracle policies are trained you can run viper on the same environment:
+
+```
+python main.py train-viper --env-name CartPole-v1 --n-env 1
+```

gym_env/__init__.py

@@ -0,0 +1,34 @@
+import gym
+import numpy as np
+from gym import register
+from stable_baselines3.common.env_util import make_atari_env
+from stable_baselines3.common.vec_env import VecFrameStack, DummyVecEnv
+
+register(
+    id='ToyPong-v0',
+    entry_point='gym_env.toy_pong:ToyPong',
+    kwargs={'args': None}
+)
+
+register(
+    id='WrappedPong-v0',
+    entry_point='gym_env.atari_pong:AtariPong',
+    kwargs={'args': None}
+)
+
+
+def make_env(args):
+    if args.env_name == "PongNoFrameskip-v4":
+        env = make_atari_env(args.env_name, n_envs=args.n_env)
+        env = VecFrameStack(env, n_stack=4)
+        return env
+    if args.env_name == "CartPole-v1":
+        return DummyVecEnv([lambda: gym.make(args.env_name) for _ in range(args.n_env)])
+    return gym.make(args.env_name)
+
+
+def is_done(done):
+    if type(done) is np.ndarray:
+        return done.all()
+    else:
+        return done

gym_env/atari_pong.py

@@ -0,0 +1,107 @@
+import collections
+import math
+
+import numpy as np
+import gym
+
+# The default ALE environment returns only the picture of the game screen,
+# however the extracted policy requires a state vector (x, y, y', y'', ball_x, ball_y)
+# that is extracted from the image, but it is not clear how.
+# We therefore extract everything we can from the ALE ram and estimate the missing quantities.
+# See https://github.com/mila-iqia/atari-representation-learning/blob/master/atariari/benchmark/ram_annotations.py#L178
+class AtariPong(gym.Env):
+    def __init__(self, args):
+        self.env = gym.make("Pong-v4", obs_type="ram", render_mode="human")
+
+        self._last_states = collections.deque(maxlen=10)
+        self.observation_space = gym.spaces.Box(low=-np.inf, high=np.inf, shape=(8,), dtype=np.float32)
+        self.action_space = self.env.action_space
+
+    def reset(self):
+        self._last_states.clear()
+        ram = self.env.reset()
+        return self._ram_to_obs(ram)
+
+    def _ram_to_obs(self, ram):
+        player_paddle_y = ram[51].astype(np.float32)  # Y coordinate of your paddle
+        ball_x = ram[49].astype(np.float32)  # X coordinate of ball
+        ball_y = ram[54].astype(np.float32)  # Y coordinate of ball
+
+        speed = self._estimate_speed()[-1]
+        acc = self._estimate_acceleration()[-1]
+        jerk = self._estimate_jerk()[-1]
+        ball_speed = self._estimate_ball_velocity()[-1]
+
+        state = [player_paddle_y, speed, acc, jerk, ball_x, ball_y, ball_speed[0], ball_speed[1]]
+
+        return state
+
+    def step(self, action):
+        ram, reward, done, info = self.env.step(action)
+
+        next_state = self._ram_to_obs(ram)
+        self._last_states.append(next_state)
+
+        return next_state, reward, done, info
+
+    def render(self, mode="human"):
+        self.env.render(mode="rgb_array")
+
+    def _estimate_speed(self):
+        if len(self._last_states) < 2:
+            return [0]
+
+        last = None
+        speeds = []
+        for state in self._last_states:
+            paddle_y = state[0]
+            if last is None:
+                last = paddle_y
+                continue
+
+            speeds.append(paddle_y - last)
+            last = paddle_y
+
+        return speeds
+
+    def _estimate_acceleration(self):
+        speeds = self._estimate_speed()
+        if len(speeds) < 2:
+            return [0]
+
+        last = None
+        accs = []
+        for speed in speeds:
+            if last is None:
+                last = speed
+                continue
+
+            accs.append(speed - last)
+            last = speed
+
+        return accs
+
+    def _estimate_jerk(self):
+        accs = self._estimate_acceleration()
+        if len(accs) < 2:
+            return [0]
+
+        return [accs[-1] - accs[-2]]
+
+    def _estimate_ball_velocity(self):
+        if len(self._last_states) < 2:
+            return [[0, 0]]
+
+        last = None
+        speeds = []
+        for state in self._last_states:
+            ball_x = state[4]
+            ball_y = state[5]
+            if last is None:
+                last = [ball_x, ball_y]
+                continue
+
+            speeds.append([ball_x - last[0], ball_y - last[1]])
+            last = [ball_x, ball_y]
+
+        return speeds

gym_env/toy_pong.py

@@ -0,0 +1,10 @@
+import gym
+import numpy as np
+from gym.spaces import Box
+
+# https://stable-baselines3.readthedocs.io/en/master/guide/rl_tips.html#tips-and-tricks-when-creating-a-custom-environment
+class ToyPong(gym.Env):
+    n_agent = 1
+
+    def __init__(self, args):
+        self.args = args