success_threshold: float = 30.0
experiment_high_reward_selfplay: TicTacToeExperiment = TicTacToeExperiment(
"experiment_high_reward", success_threshold,
environment_selfplay_high_reward, dddqn_curriculum_agent_second,
interface_high_reward_selfplay)
# Define experiments data
testing_episodes: int = 100
test_cycles: int = 10
training_episodes: int = 1000
validation_episodes: int = 100
max_training_episodes: int = 25000
episode_length_max: int = 20
# Run curriculum experiments for low reward
saved_metagraph_paths: [] = run_experiment(
experiment_low_reward_fixed, training_episodes, max_training_episodes,
episode_length_max, validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation, render_during_test,
workspace_path, __file__, logger, None, experiment_iterations_number)
for metagraph_path in saved_metagraph_paths:
run_experiment(experiment_low_reward_selfplay, training_episodes,
max_training_episodes, episode_length_max,
validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation, True,
workspace_path, __file__, logger, metagraph_path)
# Run curriculum experiments for high reward
saved_metagraph_paths: [] = run_experiment(
experiment_high_reward_fixed, training_episodes, max_training_episodes,
episode_length_max, validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation, render_during_test,
workspace_path, __file__, logger, None, experiment_iterations_number)
for metagraph_path in saved_metagraph_paths:
testing_episodes: int = 100
test_cycles: int = 10
training_episodes: int = 100
validation_episodes: int = 100
max_training_episodes: int = 35000
episode_length_max: int = 100
plot_sample_density: int = 10
# Run experiment
intro: str = "Data:\n" \
"\nVanilla Policy Gradient with GAE buffer" \
"\nThree dense layer with 4096 neurons each using xavier initialization" \
"\nLearning rate policy: 0.0003" \
"\nLearning rate advantage: 0.0001" \
"\nDiscount factor: 0.99" \
"\nValue steps per update: 10" \
"\nLambda parameter: 0.95" \
"\nUpdates per training volley: 2" \
"\nSuccess threshold: 0.35 average total reward on the validation set episodes" \
"\nEpisodic: yes" \
"\nEpisode length: 100" \
"\nMax allowed steps for episode: 100" \
"\nSeed states range [0, 0]" \
"\nAcceptance value: none" \
"\nThreshold value: 0.2\n"
run_experiment(experiment, training_episodes, max_training_episodes,
episode_length_max, validation_episodes, testing_episodes,
test_cycles, render_during_training,
render_during_validation, render_during_test,
workspace_path, __file__, logger, None,
experiment_iterations_number, intro, plot_sample_density)
environment_high_reward, dddqn_epsilon_greedy_agent,
interface_high_reward)
experiment_boltzmann_high_reward: TicTacToeExperiment = TicTacToeExperiment(
"b_experiment_high_reward", success_threshold, environment_high_reward,
dddqn_boltzmann_agent, interface_high_reward)
# Define experiments data
testing_episodes: int = 100
test_cycles: int = 10
training_episodes: int = 1000
validation_episodes: int = 100
max_training_episodes: int = 50000
episode_length_max: int = 20
# Run epsilon greedy experiment for low reward
run_experiment(experiment_egreedy_low_reward, training_episodes,
max_training_episodes, episode_length_max,
validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation,
render_during_test, workspace_path, __file__, logger, None,
experiment_iterations_number)
# Run boltzmann experiment for low reward
run_experiment(experiment_boltzmann_low_reward, training_episodes,
max_training_episodes, episode_length_max,
validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation,
render_during_test, workspace_path, __file__, logger, None,
experiment_iterations_number)
# # Run epsilon greedy experiment for high reward
run_experiment(experiment_egreedy_high_reward, training_episodes,
max_training_episodes, episode_length_max,
validation_episodes, testing_episodes, test_cycles,
render_during_training, render_during_validation,
render_during_test, workspace_path, __file__, logger, None,