Function play_game

rl/monte_carlo.py:17–55

(grid, policy, max_steps=20)

# NOTE: this is only policy evaluation, not optimization

def play_game(grid, policy, max_steps=20):
  # returns a list of states and corresponding returns

  # reset game to start at a random position
  # we need to do this, because given our current deterministic policy
  # we would never end up at certain states, but we still want to measure their value
  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()

  # keep track of all states and rewards encountered
  states = [s]
  rewards = [0]

  steps = 0
  while not grid.game_over():
    a = policy[s]
    r = grid.move(a)
    next_s = grid.current_state()

    # update states and rewards lists
    states.append(next_s)
    rewards.append(r)

    steps += 1
    if steps >= max_steps:
      break

    # update state
    # note: there is no need to store the final terminal state
    s = next_s

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

  return states, rewards


if __name__ == '__main__':