def test_reorganize(self):
trajectories = reorganize([[[1, 2], 1, [4, 5]]], [1])
self.assertEqual(np.array(trajectories).shape, (1, 1, 5))
def run(self,
is_training: bool = False,
seed: int = None) -> (List[List[dict]], dict):
"""
Run a complete game, either for evaluation or training RL agent.
:param is_training: (boolean): True if for training purpose.
:param seed: (int): A seed for running the game. For single-process program,
the seed should be set to None. For multi-process program, the
seed should be asigned for reproducibility.
:return: (tuple) Tuple containing:
(list): A list of trajectories generated from the environment.
(list): A list payoffs. Each entry corresponds to one player.
Note: The trajectories are 3-dimension list. The first dimension is for different players.
The second dimension is for different transitions. The third dimension is for the contents of each transiton
"""
if self.single_agent_mode or self.human_mode:
raise ValueError(
'Run in single agent mode or human mode is not allowed.')
if seed is not None:
np.random.seed(seed)
random.seed(seed)
trajectories = [[] for _ in range(self.player_num)]
state, player_id = self.init_game()
# Loop to play the game
trajectories[player_id].append(state)
while not self.is_over():
# Agent plays
if not is_training:
action = self.agents[player_id].eval_step(state)
else:
action = self.agents[player_id].step(state)
# Environment steps
next_state, next_player_id = self.step(action)
# Save action
trajectories[player_id].append(action)
# Set the state and player
state = next_state
player_id = next_player_id
# Save state.
if not self.game.is_over():
trajectories[player_id].append(state)
# Add a final state to all the players
for player_id in range(self.player_num):
state = self.get_state(player_id)
trajectories[player_id].append(state)
# Payoffs
payoffs = self.get_payoffs()
# Reorganize the trajectories
trajectories = reorganize(trajectories, payoffs)
return trajectories, payoffs