README.md

Random agent with action_space value checking

environment.py

import gym
import gym.scoreboard.scoring

from random_agent import RandomAgent

ENVS = ["PredictActionsCartpole-v0",
        "PredictObsCartpole-v0",
        "OffSwitchCartpole-v0",
        "SemisuperPendulumNoise-v0",
        "SemisuperPendulumRandom-v0",
        "SemisuperPendulumDecay-v0"]


def main():
    for env_name in ENVS:
        env = gym.make(env_name)

        agent = RandomAgent(env.action_space, env.observation_space)

        out_dir = '/tmp/' + agent.name + '-results'
        env.monitor.start(out_dir, force=True, video_callable=False)

        n_episodes = 1000
        for i_episode in range(n_episodes):

            observation = env.reset()
            reward = 0
            done = False

            action = agent.act(observation, reward, done)

            while not done:
                observation, reward, done, info = env.step(action)
                action = agent.act(observation, reward, done)

            print gym.scoreboard.scoring.score_from_local(out_dir)

        env.monitor.close()
        gym.upload(out_dir, algorithm_id=agent.alg_id, api_key="YOUR_API_KEY")


if __name__ == '__main__':
    main()

random_agent.py

import numpy as np


class RandomAgent(object):
    def __init__(self, action_space, observation_space):
        self.action_space = action_space
        self.observation_space = observation_space
        self.name = 'random'
        self.alg_id = "alg_MhPaN5c4TJOFS4tVFh8x3A"

    def act(self, observation, reward, done):
        return self.__validate_action(self.action_space.sample())

    # Checks for nan's in action_space samples and replaces with random value selected from a normal distribution(0, 1)
    def __validate_action(self, action):
        if hasattr(action, '__iter__'):
            for i in range(len(action)):
                self.__validate_action(action[i])
        elif np.isnan(action):
            action = np.random.normal(0, 1.0)

        return action