Function play_one

rl2/cartpole/td_lambda.py:85–118

(model, env, eps, gamma, lambda_)

# returns a list of states_and_rewards, and the total reward
def play_one(model, env, eps, gamma, lambda_):
  observation = env.reset()
  done = False
  totalreward = 0
  states_actions_rewards = []
  iters = 0
  model.reset()
  while not done and iters < 1000000:
    action = model.sample_action(observation, eps)
    prev_observation = observation
    observation, reward, done, info = env.step(action)

    if done:
      reward = -300

    # update the model
    next = model.predict(observation)
    assert(next.shape == (1, env.action_space.n))
    G = reward + gamma*np.max(next[0])
    model.update(prev_observation, action, G, gamma, lambda_)

    states_actions_rewards.append((prev_observation, action, reward))

    if reward == 1: # if we changed the reward to -200
      totalreward += reward

    iters += 1

    # if iters > 0 and iters % 1000 == 0:
    #   print(iters)
    # if done:
    #   print "finished in < 1000 steps!"

  return states_actions_rewards, totalreward


if __name__ == '__main__':