Function play_one_td

rl2/cartpole/pg_tf.py:162–189

(env, pmodel, vmodel, gamma)

def play_one_td(env, pmodel, vmodel, gamma):
  observation = env.reset()
  done = False
  totalreward = 0
  iters = 0

  while not done and iters < 2000:
    # if we reach 2000, just quit, don't want this going forever
    # the 200 limit seems a bit early
    action = pmodel.sample_action(observation)
    prev_observation = observation
    observation, reward, done, info = env.step(action)

    # if done:
    #   reward = -200

    # update the models
    V_next = vmodel.predict(observation)[0]
    G = reward + gamma*V_next
    advantage = G - vmodel.predict(prev_observation)
    pmodel.partial_fit(prev_observation, action, advantage)
    vmodel.partial_fit(prev_observation, G)

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

  return totalreward