Function play_game

rl/monte_carlo_es.py:21–55

(grid, policy, max_steps=20)

def play_game(grid, policy, max_steps=20):
  # reset game to start at a random position
  # we need to do this if we have a deterministic policy
  # we would never end up at certain states, but we still want to measure their value
  # this is called the "exploring starts" method
  start_states = list(grid.actions.keys())
  start_idx = np.random.choice(len(start_states))
  grid.set_state(start_states[start_idx])

  s = grid.current_state()
  a = np.random.choice(ALL_POSSIBLE_ACTIONS) # first action is uniformly random

  states = [s]
  actions = [a]
  rewards = [0]

  for _ in range(max_steps):
    r = grid.move(a)
    s = grid.current_state()

    rewards.append(r)
    states.append(s)
    
    if grid.game_over():
      break
    else:
      a = policy[s]
      actions.append(a)

  # we want to return:
  # states  = [s(0), s(1), ..., s(T-1), s(T)]
  # actions = [a(0), a(1), ..., a(T-1),     ]
  # rewards = [   0, R(1), ..., R(T-1), R(T)]

  return states, actions, rewards


def max_dict(d):