def _init_evaluator():
rllib_config, stop_config = get_rllib_config(seeds=get_random_seeds(1))
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name="testing_amTFT", )
evaluator.define_the_experiment_to_run(evaluation_config=rllib_config,
stop_config=stop_config,
TrainerClass=PGTrainer)
return evaluator
def _evaluate_self_and_cross_perf(
rllib_hp,
rllib_config_eval,
policies_to_load,
trainable_class,
stop,
env_config,
tune_analysis_per_exp,
):
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name=rllib_hp["exp_name"],
local_mode=rllib_hp["debug"],
use_wandb=not rllib_hp["debug"],
)
analysis_metrics_per_mode = evaluator.perform_evaluation_or_load_data(
evaluation_config=rllib_config_eval,
stop_config=stop,
policies_to_load_from_checkpoint=policies_to_load,
tune_analysis_per_exp=tune_analysis_per_exp,
TuneTrainerClass=trainable_class,
n_cross_play_per_checkpoint=min(5, rllib_hp["train_n_replicates"] - 1),
to_load_path=rllib_hp["load_plot_data"],
)
if issubclass(
rllib_hp["env_class"],
matrix_sequential_social_dilemma.MatrixSequentialSocialDilemma,
):
background_area_coord = rllib_hp["env_class"].PAYOUT_MATRIX
else:
background_area_coord = None
plot_config = PlotConfig(
xlim=rllib_hp["x_limits"],
ylim=rllib_hp["y_limits"],
markersize=5,
jitter=rllib_hp["jitter"],
xlabel="player 1 payoffs",
ylabel="player 2 payoffs",
plot_max_n_points=rllib_hp["train_n_replicates"],
x_scale_multiplier=rllib_hp["scale_multipliers"][0],
y_scale_multiplier=rllib_hp["scale_multipliers"][1],
background_area_coord=background_area_coord,
)
evaluator.plot_results(
analysis_metrics_per_mode,
plot_config=plot_config,
x_axis_metric=f"policy_reward_mean/{env_config['players_ids'][0]}",
y_axis_metric=f"policy_reward_mean/{env_config['players_ids'][1]}",
)
def _init_evaluator():
exp_name, _ = log.log_in_current_day_dir("testing")
rllib_config, stop_config = get_rllib_config(seeds=get_random_seeds(1))
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name=exp_name,
)
evaluator.define_the_experiment_to_run(
evaluation_config=rllib_config,
stop_config=stop_config,
TrainerClass=PGTrainer,
)
return evaluator
def evaluate_self_play_cross_play(tune_analysis_per_welfare, config_eval,
env_config, stop, hp_eval):
exp_name = os.path.join(hp_eval["exp_name"], "eval")
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name=exp_name,
local_mode=hp_eval["debug"],
)
analysis_metrics_per_mode = evaluator.perform_evaluation_or_load_data(
evaluation_config=config_eval,
stop_config=stop,
policies_to_load_from_checkpoint=copy.deepcopy(
env_config["players_ids"]),
tune_analysis_per_exp=tune_analysis_per_welfare,
TrainerClass=dqn.DQNTrainer,
n_self_play_per_checkpoint=hp_eval["n_self_play_per_checkpoint"],
n_cross_play_per_checkpoint=hp_eval["n_cross_play_per_checkpoint"],
to_load_path=hp_eval["load_plot_data"],
)
if "CoinGame" in hp_eval["env_name"]:
background_area_coord = None
else:
background_area_coord = hp_eval["env_class"].PAYOUT_MATRIX
plot_config = plot.PlotConfig(
xlim=hp_eval["x_limits"],
ylim=hp_eval["y_limits"],
markersize=5,
alpha=1.0,
jitter=hp_eval["jitter"],
xlabel="player 1 payoffs",
ylabel="player 2 payoffs",
plot_max_n_points=hp_eval["train_n_replicates"],
x_scale_multiplier=hp_eval["plot_axis_scale_multipliers"][0],
y_scale_multiplier=hp_eval["plot_axis_scale_multipliers"][1],
background_area_coord=background_area_coord,
)
evaluator.plot_results(
analysis_metrics_per_mode,
plot_config=plot_config,
x_axis_metric=f"policy_reward_mean/{env_config['players_ids'][0]}",
y_axis_metric=f"policy_reward_mean/{env_config['players_ids'][1]}",
)
print_inequity_aversion_welfare(env_config, analysis_metrics_per_mode)
return analysis_metrics_per_mode
def evaluate_self_and_cross_perf(tune_analysis_per_welfare, hp):
config_eval, env_config, stop, hp_eval = generate_eval_config(hp)
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name=hp_eval["exp_name"])
analysis_metrics_per_mode = evaluator.perform_evaluation_or_load_data(
evaluation_config=config_eval,
stop_config=stop,
policies_to_load_from_checkpoint=copy.deepcopy(
env_config["players_ids"]),
tune_analysis_per_exp=tune_analysis_per_welfare,
TrainerClass=dqn.DQNTrainer,
n_cross_play_per_checkpoint=min(
5,
(hp_eval["train_n_replicates"] * len(hp_eval["welfare_functions"]))
- 1),
to_load_path=hp_eval["load_plot_data"])
if hp["env"] in [coin_game.CoinGame, coin_game.AsymCoinGame]:
background_area_coord = None
else:
background_area_coord = hp['env'].PAYOUT_MATRIX
plot_config = PlotConfig(
xlim=hp_eval["x_limits"],
ylim=hp_eval["y_limits"],
markersize=5,
alpha=1.0,
jitter=hp_eval["jitter"],
xlabel="player 1 payoffs",
ylabel="player 2 payoffs",
plot_max_n_points=hp_eval["train_n_replicates"],
# title="cross and same-play performances: " + hp_eval['env'].NAME,
x_scale_multiplier=hp_eval["plot_axis_scale_multipliers"][0],
y_scale_multiplier=hp_eval["plot_axis_scale_multipliers"][1],
background_area_coord=background_area_coord)
evaluator.plot_results(
analysis_metrics_per_mode,
plot_config=plot_config,
x_axis_metric=f"policy_reward_mean/{env_config['players_ids'][0]}",
y_axis_metric=f"policy_reward_mean/{env_config['players_ids'][1]}")
return analysis_metrics_per_mode
def evaluate_self_and_cross_perf(rllib_hp, rllib_config_eval, policies_to_load,
trainable_class, stop, env_config,
tune_analysis_per_exp):
evaluator = self_and_cross_perf.SelfAndCrossPlayEvaluator(
exp_name=rllib_hp["exp_name"])
analysis_metrics_per_mode = evaluator.perform_evaluation_or_load_data(
evaluation_config=rllib_config_eval,
stop_config=stop,
policies_to_load_from_checkpoint=policies_to_load,
tune_analysis_per_exp=tune_analysis_per_exp,
TuneTrainerClass=trainable_class,
n_cross_play_per_checkpoint=min(5, rllib_hp["train_n_replicates"] - 1),
to_load_path=rllib_hp["load_plot_data"])
if trainable_class == LOLAPGMatrice:
background_area_coord = rllib_hp['env'].PAYOUT_MATRIX
else:
background_area_coord = None
plot_config = PlotConfig(
xlim=rllib_hp["x_limits"],
ylim=rllib_hp["y_limits"],
markersize=5,
alpha=1.0,
jitter=rllib_hp["jitter"],
xlabel="player 1 payoffs",
ylabel="player 2 payoffs",
plot_max_n_points=rllib_hp["train_n_replicates"],
# title="cross and same-play performances: " + rllib_hp['env'].NAME,
x_scale_multiplier=rllib_hp["scale_multipliers"][0],
y_scale_multiplier=rllib_hp["scale_multipliers"][1],
background_area_coord=background_area_coord)
evaluator.plot_results(
analysis_metrics_per_mode,
plot_config=plot_config,
x_axis_metric=f"policy_reward_mean/{env_config['players_ids'][0]}",
y_axis_metric=f"policy_reward_mean/{env_config['players_ids'][1]}")