Random agent for Cart Pole environment

random_cart_pole_agent.py

import os
import logging
import tempfile
import numpy as np

import gym
#from gym.wrappers.monitoring import Monitor

class RandomAgent(object):
    def __init__ (self, action_space):
        self.action_space = action_space
        assert isinstance(action_space, gym.spaces.discrete.Discrete), 'unsupported action space for now.'
    
    def act(self, current_observation, last_reward, done):
        a = self.action_space.sample()
        return a


if __name__ == '__main__':
  
  # logging  
  logger = logging.getLogger()
  logger.setLevel(logging.INFO)

  # create environment object
  env = gym.make('CartPole-v0')
  # create agent 
  agent = RandomAgent(env.action_space)
  # define maximum number of episodes of the game to be played
  max_episodes = 500
  # define maximum number of steps per episode
  max_steps_per_episode = 200
  
#  Monitor(env, outdir + '/' + ENV_NAME, force = True)  
  # Untill we are not in a terminal state
    # create action object
    # sample action space and store the sample in the action object
    # pass this object to environment
    # collect the reward, state from environment
    # render the updated state
  sum_reward_running = 0.0  
  current_reward = 0
  done = False

  for i in xrange(max_episodes):
      current_observation = env.reset() # starting fresh episode
      sum_rewards = 0 # reset rewards for next episode
      last_reward = 0
      for j in xrange(max_steps_per_episode):
          action = agent.act(current_observation, last_reward, done)
          next_observation, current_reward, done, _ = env.step(action)
          sum_rewards += current_reward
          if done:
              break
          current_observation = next_observation # if not done, go for next iteration in the current episode
          last_reward = current_reward

      sum_reward_running = 0.95 * sum_reward_running + sum_rewards * 0.05 # cumulative reward across all episodes of the game play
      print '%d running reward: %f' % (i, sum_reward_running)
  # just done playing all max_episode game sessions!!
  # dump monitor info to disk
 # env.monitor.close()