def main(argv: Sequence[str]) -> None:
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
param_names, param_values = zip(
*[convert_param_spec(spec) for spec in FLAGS.parameters])
header = (['game_name'] + list(param_names) +
['fictitious_play_iteration_time'])
timing_results = []
for game_name in FLAGS.games:
for param_tuple in itertools.product(*param_values):
result_line = [game_name] + [str(p) for p in param_tuple]
print('Computing timings for:', ' '.join(result_line))
param_dict = dict(zip(param_names, param_tuple))
game = pyspiel.load_game(game_name, param_dict)
t0 = time.time()
fp = fictitious_play.FictitiousPlay(game)
fp.iteration()
elapsed = time.time() - t0
result_line.append(f'{elapsed:.4f}s')
print(' '.join(result_line))
timing_results.append(result_line)
print('\nRESULTS:')
print(' '.join(header))
for line in timing_results:
print(' '.join([str(v) for v in line]))
def test_dqn_fp_python_game(self):
"""Checks if fictitious play with DQN-based value function works."""
game = crowd_modelling.MFGCrowdModellingGame()
dfp = fictitious_play.FictitiousPlay(game)
uniform_policy = policy.UniformRandomPolicy(game)
dist = distribution.DistributionPolicy(game, uniform_policy)
envs = [
rl_environment.Environment(game,
mfg_distribution=dist,
mfg_population=p)
for p in range(game.num_players())
]
dqn_agent = dqn.DQN(
0,
state_representation_size=envs[0].observation_spec()["info_state"]
[0],
num_actions=envs[0].action_spec()["num_actions"],
hidden_layers_sizes=[256, 128, 64],
replay_buffer_capacity=100,
batch_size=5,
epsilon_start=0.02,
epsilon_end=0.01)
for _ in range(10):
dfp.iteration(rl_br_agent=dqn_agent)
dfp_policy = dfp.get_policy()
nash_conv_dfp = nash_conv.NashConv(game, dfp_policy)
self.assertAlmostEqual(nash_conv_dfp.nash_conv(), 1.0558451955622807)
def test_fp_cpp_game(self):
"""Checks if fictitious play works."""
game = pyspiel.load_game("mfg_crowd_modelling")
fp = fictitious_play.FictitiousPlay(game)
for _ in range(10):
fp.iteration()
fp_policy = fp.get_policy()
nash_conv_fp = nash_conv.NashConv(game, fp_policy)
self.assertAlmostEqual(nash_conv_fp.nash_conv(), 0.9908032626911343)
def test_fp_python_game(self):
"""Checks if fictitious play works."""
game = crowd_modelling.MFGCrowdModellingGame()
fp = fictitious_play.FictitiousPlay(game)
for _ in range(10):
fp.iteration()
fp_policy = fp.get_policy()
nash_conv_fp = nash_conv.NashConv(game, fp_policy)
self.assertAlmostEqual(nash_conv_fp.nash_conv(), 0.9908032626911343)
def mean_field_fictitious_play(mfg_game,
number_of_iterations,
compute_metrics=False):
fp = mean_field_fictitious_play_module.FictitiousPlay(mfg_game)
tick_time = time.time()
for _ in range(number_of_iterations):
fp.iteration()
timing = time.time() - tick_time
fp_policy = fp.get_policy()
# print('learning done')
if compute_metrics:
distribution_mfg = distribution_module.DistributionPolicy(
mfg_game, fp_policy)
# print('distribution done')
policy_value_ = policy_value.PolicyValue(
mfg_game, distribution_mfg,
fp_policy).value(mfg_game.new_initial_state())
nash_conv_fp = nash_conv_module.NashConv(mfg_game, fp_policy)
return timing, fp_policy, nash_conv_fp, policy_value_
return timing, fp_policy
def main(argv: Sequence[str]) -> None:
# TODO(perolat): move to an example directory.
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
mfg_game = pyspiel.load_game(FLAGS.game, GAME_SETTINGS.get(FLAGS.game, {}))
mfg_state = mfg_game.new_initial_state()
print('Playing a single arbitrary trajectory')
while not mfg_state.is_terminal():
print('State obs string:', mfg_state.observation_string(0))
if mfg_state.current_player() == pyspiel.PlayerId.CHANCE:
action_list, prob_list = zip(*mfg_state.chance_outcomes())
action = np.random.choice(action_list, p=prob_list)
mfg_state.apply_action(action)
elif mfg_state.current_player() == pyspiel.PlayerId.MEAN_FIELD:
dist_to_register = mfg_state.distribution_support()
n_states = len(dist_to_register)
dist = [1.0 / n_states for _ in range(n_states)]
mfg_state.update_distribution(dist)
else:
legal_list = mfg_state.legal_actions()
action = np.random.choice(legal_list)
mfg_state.apply_action(action)
print('compute nashconv')
uniform_policy = policy.UniformRandomPolicy(mfg_game)
nash_conv_fp = nash_conv.NashConv(mfg_game, uniform_policy)
print('Nashconv:', nash_conv_fp.nash_conv())
print('compute distribution')
mfg_dist = distribution.DistributionPolicy(mfg_game, uniform_policy)
br_value = best_response_value.BestResponse(
mfg_game, mfg_dist, value.TabularValueFunction(mfg_game))
py_value = policy_value.PolicyValue(mfg_game, mfg_dist, uniform_policy,
value.TabularValueFunction(mfg_game))
print(
'Value of a best response policy to a uniform policy '
'(computed with best_response_value)',
br_value(mfg_game.new_initial_state()))
print('Value of the uniform policy:', py_value(mfg_game.new_initial_state()))
greedy_pi = greedy_policy.GreedyPolicy(mfg_game, None, br_value)
greedy_pi = greedy_pi.to_tabular()
pybr_value = policy_value.PolicyValue(mfg_game, mfg_dist, greedy_pi,
value.TabularValueFunction(mfg_game))
print(
'Value of a best response policy to a uniform policy (computed at the '
'value of the greedy policy of the best response value)',
pybr_value(mfg_game.new_initial_state()))
print('merge')
merged_pi = fictitious_play.MergedPolicy(
mfg_game, list(range(mfg_game.num_players())),
[uniform_policy, greedy_pi],
[mfg_dist, distribution.DistributionPolicy(mfg_game, greedy_pi)],
[0.5, 0.5])
merged_pi_value = policy_value.PolicyValue(
mfg_game, mfg_dist, merged_pi, value.TabularValueFunction(mfg_game))
print(br_value(mfg_game.new_initial_state()))
print(py_value(mfg_game.new_initial_state()))
print(merged_pi_value(mfg_game.new_initial_state()))
print((br_value(mfg_game.new_initial_state()) +
py_value(mfg_game.new_initial_state())) / 2)
print('fp')
fp = fictitious_play.FictitiousPlay(mfg_game)
for j in range(100):
print('Iteration', j, 'of fictitious play')
fp.iteration()
fp_policy = fp.get_policy()
nash_conv_fp = nash_conv.NashConv(mfg_game, fp_policy)
print('Nashconv of the current FP policy', nash_conv_fp.nash_conv())
print('md')
md = mirror_descent.MirrorDescent(mfg_game,
value.TabularValueFunction(mfg_game))
for j in range(10):
print('Iteration', j, 'of mirror descent')
md.iteration()
md_policy = md.get_policy()
nash_conv_md = nash_conv.NashConv(mfg_game, md_policy)
print('Nashconv of the current MD policy', nash_conv_md.nash_conv())
def main(argv: Sequence[str]) -> None:
# TODO(perolat): move to an example directory.
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
game_settings = {
'only_distribution_reward': True,
'forbidden_states': '[0|0;0|1]',
'initial_distribution': '[0|2;0|3]',
'initial_distribution_value': '[0.5;0.5]',
}
mfg_game = pyspiel.load_game(FLAGS.game, game_settings)
mfg_state = mfg_game.new_initial_state()
while not mfg_state.is_terminal():
print(mfg_state.observation_string(0))
if mfg_state.current_player() == pyspiel.PlayerId.CHANCE:
action_list, prob_list = zip(*mfg_state.chance_outcomes())
action = np.random.choice(action_list, p=prob_list)
mfg_state.apply_action(action)
elif mfg_state.current_player() == pyspiel.PlayerId.MEAN_FIELD:
dist_to_register = mfg_state.distribution_support()
n_states = len(dist_to_register)
dist = [1.0 / n_states for _ in range(n_states)]
mfg_state.update_distribution(dist)
else:
legal_list = mfg_state.legal_actions()
action = np.random.choice(legal_list)
mfg_state.apply_action(action)
print('compute nashconv')
uniform_policy = policy.UniformRandomPolicy(mfg_game)
nash_conv_fp = nash_conv.NashConv(mfg_game, uniform_policy)
print(nash_conv_fp.nash_conv())
print('compute distribution')
mfg_dist = distribution.DistributionPolicy(mfg_game, uniform_policy)
br_value = best_response_value.BestResponse(mfg_game, mfg_dist)
py_value = policy_value.PolicyValue(mfg_game, mfg_dist, uniform_policy)
print(br_value(mfg_game.new_initial_state()))
print(py_value(mfg_game.new_initial_state()))
greedy_pi = greedy_policy.GreedyPolicy(mfg_game, None, br_value)
greedy_pi = greedy_pi.to_tabular()
pybr_value = policy_value.PolicyValue(mfg_game, mfg_dist, greedy_pi)
print(pybr_value(mfg_game.new_initial_state()))
print('merge')
merged_pi = fictitious_play.MergedPolicy(
mfg_game, list(range(mfg_game.num_players())),
[uniform_policy, greedy_pi],
[mfg_dist,
distribution.DistributionPolicy(mfg_game, greedy_pi)], [0.5, 0.5])
merged_pi_value = policy_value.PolicyValue(mfg_game, mfg_dist, merged_pi)
print(br_value(mfg_game.new_initial_state()))
print(py_value(mfg_game.new_initial_state()))
print(merged_pi_value(mfg_game.new_initial_state()))
print((br_value(mfg_game.new_initial_state()) +
py_value(mfg_game.new_initial_state())) / 2)
print('fp')
fp = fictitious_play.FictitiousPlay(mfg_game)
for j in range(100):
print(j)
fp.iteration()
fp_policy = fp.get_policy()
nash_conv_fp = nash_conv.NashConv(mfg_game, fp_policy)
print(nash_conv_fp.nash_conv())
