def train_sarsa_models():
params = {"seed": None}
for seed in get_seeds():
params["seed"] = seed
train_experiment(get_sarsa_experiment(params), get_states(),
get_rewards())
def aggregate_models_by_avg(experiment_getter, seeds, params_getter, state_getter, reward_getter):
aggregated_models = {}
for seed in seeds:
experiment = experiment_getter(params_getter(seed))
models, states, rewards, filenames = get_models_from_experiment(experiment,
state_getter(),
reward_getter(),
train_if_missing=True)
for model, state, reward in zip(models, states, rewards):
key = (state, reward)
if key not in aggregated_models:
aggregated_models[key] = model
else:
aggregated_models[key].food_count_per_episode += model.food_count_per_episode
aggregated_models[key].rewards_per_episode += model.rewards_per_episode
for seed in get_seeds()[1:]:
params = params_getter(seed)
tmp_experiment = experiment_getter(params)
tmp_models, tmp_states, tmp_rewards, _ = get_models_from_experiment(
tmp_experiment,
state_getter(),
reward_getter(),
train_if_missing=True)
for true_model, true_state in zip(models, states):
for tmp_model, tmp_state in zip(models, states):
if true_state == tmp_state:
true_model.food_count_per_episode += tmp_model.food_count_per_episode
true_model.rewards_per_episode += tmp_model.rewards_per_episode
for _, model in aggregated_models.items():
model.food_count_per_episode = model.food_count_per_episode / len(seeds)
model.rewards_per_episode = model.rewards_per_episode / len(seeds)
models, states, rewards = [], [], []
for (state, reward), model in aggregated_models.items():
models.append(model)
states.append(state)
rewards.append(reward)
return models, states, rewards
def experiment01_sarsa():
def get_params(seed):
image_output = os.path.join(get_project_path(),
"images",
"sarsa",
"state",
str(seed),
"state_sarsa_average_game_score_over_time.png")
return {"seed": seed,
"image_output": image_output}
params = get_params(42)
experiment = get_sarsa_experiment(params)
models, states, rewards = aggregate_models_by_avg(get_sarsa_experiment, get_seeds(), get_params, get_states(),
get_rewards())
experiment01(experiment, models, states, rewards, params)
def experiment01_sarsa_directional_state():
def get_params(seed):
image_output = os.path.join(
get_project_path(), "images", "sarsa", "reward", str(seed),
"reward_sarsa_directional_state_average_game_score_over_time.png")
return {"seed": seed, "image_output": image_output}
params = get_params(42)
experiment = get_qlearning_experiment(params)
models, states, rewards = aggregate_models_by_avg(get_sarsa_experiment,
get_seeds(), get_params,
get_states, get_rewards)
models, states, rewards = \
filter_models_with_rewards_by_state(models, states, rewards,
lambda state:
state.startswith("Directional"))
experiment01(experiment, models, states, rewards, params)
def get_experiment(): experiment = setup_experiment(QLearning, "info_augmentation", get_seeds()[0]) experiment.train_episodes = 1000000 return experiment