OpenAI GymのPendulum-v0をDDPGで解く

Pendulum-v0(回転倒立振子)をKeras-RL¹のDDPG²で解いてみました．

問題³

• https://gym.openai.com/envs/Pendulum-v0
• https://github.com/openai/gym/wiki/Pendulum-v0

状態空間

Num	Observation	Min	Max
0	$\cos(\theta)$ (x座標)	-1	1
1	$\sin(\theta)$ (y座標)	-1	1
2	$\dot{\theta}$ (角速度)	-8	8

行動空間

Num	Action	Min	Max
0	$a$ (Joint effort)	-2	2

報酬

$R_t = -\left ( \theta^2 + 0.1 \dot{\theta}^2 + 0.001 a^2 \right )$

DDPGとは

解法

コード

import numpy as np
import gym
from gym import wrappers
from keras.models import Model
from keras.layers import Dense, Flatten, Input, concatenate
from keras.optimizers import Adam
from rl.agents import DDPGAgent
from rl.memory import SequentialMemory

def build_actor_model(num_action, observation_shape):
    action_input = Input(shape=(1,)+observation_shape)
    x = Flatten()(action_input)
    x = Dense(16, activation="relu")(x)
    x = Dense(16, activation="relu")(x)
    x = Dense(num_action, activation="linear")(x)
    actor = Model(inputs=action_input, outputs=x)
    return actor

def build_critic_model(num_action, observation_shape):
    action_input = Input(shape=(num_action,))
    observation_input = Input(shape=(1,)+observation_shape)
    flattened_observation = Flatten()(observation_input)
    x = concatenate([action_input, flattened_observation])
    x = Dense(32, activation="relu")(x)
    x = Dense(32, activation="relu")(x)
    x = Dense(1, activation="linear")(x)
    critic = Model(inputs=[action_input, observation_input], outputs=x)
    return (critic, action_input)

def build_agent(num_action, observation_shape):
    actor = build_actor_model(num_action, observation_shape)
    critic, critic_action_input = build_critic_model(num_action, observation_shape)
    memory = SequentialMemory(limit=10**5, window_length=1)
    agent = DDPGAgent(
        num_action,
        actor,
        critic,
        critic_action_input,
        memory
    )
    return agent

def run():
    env = gym.make("Pendulum-v0")
    env = wrappers.Monitor(env, directory="/tmp/pendulum-v0", force=True)
    print("Action Space: %s" % env.action_space)
    print("Observation Space: %s" % env.observation_space)
    agent = build_agent(env.action_space.shape[0], env.observation_space.shape)
    agent.compile(Adam(lr=0.001, clipnorm=1.), metrics=["mae"])
    agent.fit(env, nb_steps=50000, visualize=True, verbose=1, nb_max_episode_steps=200)
    agent.test(env, nb_episodes=5, visualize=True, nb_max_episode_steps=200)

if __name__ == "__main__":
    run()

スコア

-382.02 ± 48.13

References

1. GitHub - Keras-RL, Plappert et al., 2016. ↩

2. Continuous control with deep reinforcement learning, Lillicrap et al., 2015. ↩

3. GitHub Wiki - Pendulumn v0 ↩