config.hyperparameters = {
"HRL": {
"linear_hidden_units":
linear_hidden_units,
"learning_rate":
learning_rate,
"buffer_size":
buffer_size,
"batch_size":
batch_size,
"final_layer_activation":
"None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions":
[[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm":
batch_norm,
"gradient_clipping_norm":
gradient_clipping_norm,
"update_every_n_steps":
update_every_n_steps,
"epsilon_decay_rate_denominator":
epsilon_decay_rate_denominator,
"discount_rate":
discount_rate,
"learning_iterations":
learning_iterations,
"tau":
tau,
"sequitur_k":
sequitur_k,
"action_length_reward_bonus":
action_length_reward_bonus,
"pre_training_learning_iterations_multiplier":
pre_training_learning_iterations_multiplier,
"episodes_to_run_with_no_exploration":
episodes_to_run_with_no_exploration,
"action_balanced_replay_buffer":
action_balanced_replay_buffer,
"copy_over_hidden_layers":
copy_over_hidden_layers
},
"DQN_Agents": {
"linear_hidden_units":
linear_hidden_units,
"learning_rate":
learning_rate,
"buffer_size":
buffer_size,
"batch_size":
batch_size,
"final_layer_activation":
"None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions":
[[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm":
batch_norm,
"gradient_clipping_norm":
gradient_clipping_norm,
"update_every_n_steps":
update_every_n_steps,
"epsilon_decay_rate_denominator":
epsilon_decay_rate_denominator,
"discount_rate":
discount_rate,
"learning_iterations":
learning_iterations,
"tau":
tau,
},
"Actor_Critic_Agents": {
"Actor": {
"learning_rate":
0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation":
"Softmax",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": [[
config.environment.observation_space.n,
embedding_dimensionality
]],
"batch_norm":
False,
"tau":
0.005,
"gradient_clipping_norm":
5,
"initialiser":
"Xavier"
},
"Critic": {
"learning_rate":
0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation":
None,
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": [[
config.environment.observation_space.n,
embedding_dimensionality
]],
"batch_norm":
False,
"buffer_size":
1000000,
"tau":
0.005,
"gradient_clipping_norm":
5,
"initialiser":
"Xavier"
},
"min_steps_before_learning": 10000,
"batch_size": 256,
"discount_rate": 0.99,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 1,
"learning_updates_per_learning_session": 1,
"automatically_tune_entropy_hyperparameter": True,
"entropy_term_weight": None,
"add_extra_noise": False,
"do_evaluation_iterations": True
}
}
config.hyperparameters = {
"DQN_Agents": {
"learning_rate": 0.005,
"batch_size": 64,
"buffer_size": 40000,
"epsilon": 0.1,
"epsilon_decay_rate_denominator": 200,
"discount_rate": 0.99,
"tau": 0.1,
"alpha_prioritised_replay": 0.6,
"beta_prioritised_replay": 0.4,
"incremental_td_error": 1e-8,
"update_every_n_steps": 3,
"linear_hidden_units": [20, 20, 20],
"final_layer_activation": "None",
"batch_norm": False,
"gradient_clipping_norm": 5,
"HER_sample_proportion": 0.8,
"clip_rewards": False,
"learning_iterations": 1
},
"Stochastic_Policy_Search_Agents": {
"policy_network_type": "Linear",
"noise_scale_start": 1e-2,
"noise_scale_min": 1e-3,
"noise_scale_max": 2.0,
"noise_scale_growth_factor": 2.0,
"stochastic_action_decision": False,
"num_policies": 10,
"episodes_per_policy": 1,
"num_policies_to_keep": 5,
"clip_rewards": False
},
"Policy_Gradient_Agents": {
"learning_rate": 0.01,
"linear_hidden_units": [20],
"final_layer_activation": "SOFTMAX",
"learning_iterations_per_round": 7,
"discount_rate": 0.99,
"batch_norm": False,
"clip_epsilon": 0.1,
"episodes_per_learning_round": 7,
"normalise_rewards": False,
"gradient_clipping_norm": 5,
"mu": 0.0, #only required for continuous action games
"theta": 0.0, #only required for continuous action games
"sigma": 0.0, #only required for continuous action games
"epsilon_decay_rate_denominator": 1,
"clip_rewards": False
},
"Actor_Critic_Agents": {
"learning_rate": 0.0005,
"linear_hidden_units": [150, 30, 30, 30],
"final_layer_activation": ["SOFTMAX", None],
"gradient_clipping_norm": 25.0,
"discount_rate": 0.99,
"epsilon_decay_rate_denominator": 10.0,
"normalise_rewards": False,
"automatically_tune_entropy_hyperparameter": True,
"add_extra_noise": False,
"min_steps_before_learning": 4,
"do_evaluation_iterations": True,
"clip_rewards": False,
"Actor": {
"learning_rate": 0.001,
"linear_hidden_units": [20, 20],
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"linear_hidden_units": [20, 20],
"final_layer_activation": "None",
"batch_norm": False,
"buffer_size": 100000,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"batch_size": 3,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 20,
"learning_updates_per_learning_session": 10,
"HER_sample_proportion": 0.8,
"exploration_worker_difference": 1.0
},
"SNN_HRL": {
"SKILL_AGENT": {
"num_skills": 20,
"regularisation_weight": 1.5,
"visitations_decay": 0.9999,
"episodes_for_pretraining": 7,
"batch_size": 256,
"learning_rate": 0.001,
"buffer_size": 40000,
"linear_hidden_units": [20, 10],
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0, 1],
"embedding_dimensions": [[300, 10], [20, 6]],
"batch_norm": False,
"gradient_clipping_norm": 2,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 500,
"discount_rate": 0.999,
"learning_iterations": 1,
"tau": 0.01,
"clip_rewards": False
},
"MANAGER": {
"timesteps_before_changing_skill": 6,
"linear_hidden_units": [10, 5],
"learning_rate": 0.01,
"buffer_size": 40000,
"batch_size": 3,
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": [[300, 10]],
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 50,
"discount_rate": 0.99,
"learning_iterations": 1,
"tau": 0.01,
"clip_rewards": False
}
}
}
config.hyperparameters = {
"HRL": {
"linear_hidden_units": linear_hidden_units,
"learning_rate": learning_rate,
"buffer_size": buffer_size,
"batch_size": batch_size,
"final_layer_activation": "None",
# "columns_of_data_to_be_embedded": [0],
# "embedding_dimensions": [[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm": batch_norm,
"gradient_clipping_norm": gradient_clipping_norm,
"update_every_n_steps": update_every_n_steps,
"epsilon_decay_rate_denominator": epsilon_decay_rate_denominator,
"discount_rate": discount_rate,
"learning_iterations": learning_iterations,
"tau": tau,
"sequitur_k": sequitur_k,
"use_relative_counts": use_relative_counts,
"action_length_reward_bonus": action_length_reward_bonus,
"pre_training_learning_iterations_multiplier": pre_training_learning_iterations_multiplier,
"episodes_to_run_with_no_exploration": episodes_to_run_with_no_exploration,
"action_balanced_replay_buffer": action_balanced_replay_buffer,
"copy_over_hidden_layers": copy_over_hidden_layers,
"random_episodes_to_run": random_episodes_to_run,
"only_train_new_actions": only_train_new_actions,
"only_train_final_layer": only_train_final_layer,
"num_top_results_to_use": num_top_results_to_use,
"action_frequency_required_in_top_results": action_frequency_required_in_top_results,
"reduce_macro_action_appearance_cutoff_throughout_training": reduce_macro_action_appearance_cutoff_throughout_training,
"add_1_macro_action_at_a_time": add_1_macro_action_at_a_time,
"calculate_q_values_as_increments": calculate_q_values_as_increments,
"episodes_per_round": episodes_per_round,
"abandon_ship": abandon_ship,
"clip_rewards": clip_rewards
},
"DQN_Agents": {
"linear_hidden_units": linear_hidden_units,
"learning_rate": learning_rate,
"buffer_size": buffer_size,
"batch_size": batch_size,
"final_layer_activation": "None",
# "columns_of_data_to_be_embedded": [0],
# "embedding_dimensions": [[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm": batch_norm,
"gradient_clipping_norm": gradient_clipping_norm,
"update_every_n_steps": update_every_n_steps,
"epsilon_decay_rate_denominator": epsilon_decay_rate_denominator,
"discount_rate": discount_rate,
"learning_iterations": learning_iterations,
"tau": tau,
"clip_rewards": clip_rewards
},
"Actor_Critic_Agents": {
"Actor": {
"learning_rate": 0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation": "Softmax",
# "columns_of_data_to_be_embedded": [0],
# "embedding_dimensions": [[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"Critic": {
"learning_rate": 0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation": None,
# "columns_of_data_to_be_embedded": [0],
# "embedding_dimensions": [[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm": False,
"buffer_size": 1000000,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"min_steps_before_learning": 10000,
"batch_size": 256,
"discount_rate": 0.99,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 1,
"learning_updates_per_learning_session": 1,
"automatically_tune_entropy_hyperparameter": True,
"entropy_term_weight": None,
"add_extra_noise": False,
"do_evaluation_iterations": True,
"clip_rewards": clip_rewards
}
}
config.hyperparameters = {
"linear_hidden_units":
linear_hidden_units,
"learning_rate":
learning_rate,
"buffer_size":
buffer_size,
"batch_size":
batch_size,
"final_layer_activation":
"None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions":
[[config.environment.observation_space.n, embedding_dimensionality]],
"batch_norm":
batch_norm,
"gradient_clipping_norm":
gradient_clipping_norm,
"update_every_n_steps":
update_every_n_steps,
"epsilon_decay_rate_denominator":
epsilon_decay_rate_denominator,
"discount_rate":
discount_rate,
"learning_iterations":
learning_iterations,
"tau":
tau,
"sequitur_k":
sequitur_k,
"action_length_reward_bonus":
0.1,
"episodes_to_run_with_no_exploration":
10,
"pre_training_learning_iterations_multiplier":
0.1,
"copy_over_hidden_layers":
True,
"use_global_list_of_best_performing_actions":
True
}
config.hyperparameters = {
"DQN_Agents": {
"learning_rate": 0.01,
"batch_size": 256,
"buffer_size": 40000,
"epsilon": 1.0,
"epsilon_decay_rate_denominator": 1,
"discount_rate": 0.99,
"tau": 0.01,
"alpha_prioritised_replay": 0.6,
"beta_prioritised_replay": 0.1,
"incremental_td_error": 1e-8,
"update_every_n_steps": 1,
"linear_hidden_units": [30, 15],
"final_layer_activation": "None",
"batch_norm": False,
"gradient_clipping_norm": 0.7,
"learning_iterations": 1,
"clip_rewards": False
},
"Stochastic_Policy_Search_Agents": {
"policy_network_type": "Linear",
"noise_scale_start": 1e-2,
"noise_scale_min": 1e-3,
"noise_scale_max": 2.0,
"noise_scale_growth_factor": 2.0,
"stochastic_action_decision": False,
"num_policies": 10,
"episodes_per_policy": 1,
"num_policies_to_keep": 5,
"clip_rewards": False
},
"Policy_Gradient_Agents": {
"learning_rate": 0.05,
"linear_hidden_units": [20, 20],
"final_layer_activation": "SOFTMAX",
"learning_iterations_per_round": 5,
"discount_rate": 0.99,
"batch_norm": False,
"clip_epsilon": 0.1,
"episodes_per_learning_round": 4,
"normalise_rewards": True,
"gradient_clipping_norm": 7.0,
"mu": 0.0, #only required for continuous action games
"theta": 0.0, #only required for continuous action games
"sigma": 0.0, #only required for continuous action games
"epsilon_decay_rate_denominator": 1.0,
"clip_rewards": False
},
"Actor_Critic_Agents": {
"learning_rate": 0.005,
"linear_hidden_units": [20, 10],
"final_layer_activation": ["SOFTMAX", None],
"gradient_clipping_norm": 5.0,
"discount_rate": 0.99,
"epsilon_decay_rate_denominator": 1.0,
"normalise_rewards": True,
"exploration_worker_difference": 2.0,
"clip_rewards": False,
"Actor": {
"learning_rate": 0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation": "Softmax",
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"Critic": {
"learning_rate": 0.0003,
"linear_hidden_units": [64, 64],
"final_layer_activation": None,
"batch_norm": False,
"buffer_size": 1000000,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"min_steps_before_learning": 400,
"batch_size": 256,
"discount_rate": 0.99,
"mu": 0.0, #for O-H noise
"theta": 0.15, #for O-H noise
"sigma": 0.25, #for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 1,
"learning_updates_per_learning_session": 1,
"automatically_tune_entropy_hyperparameter": True,
"entropy_term_weight": None,
"add_extra_noise": False,
"do_evaluation_iterations": True
}
}
config.visualise_overall_agent_results = False
config.randomise_random_seed = False
config.runs_per_agent = 1
config.use_GPU = False
config.hyperparameters = {
"DQN_Agents": {
"learning_rate": 0.005,
"batch_size": 3,
"buffer_size": 40000,
"epsilon": 0.1,
"epsilon_decay_rate_denominator": 200,
"discount_rate": 0.99,
"tau": 0.1,
"alpha_prioritised_replay": 0.6,
"beta_prioritised_replay": 0.4,
"incremental_td_error": 1e-8,
"update_every_n_steps": 3,
"linear_hidden_units": [20, 20, 20],
"final_layer_activation": "None",
"batch_norm": False,
"gradient_clipping_norm": 5,
"HER_sample_proportion": 0.8,
"clip_rewards": False
}
}
trainer = Trainer(config, [DQN_HER])
config.hyperparameters = config.hyperparameters["DQN_Agents"]
config.hyperparameters = {
"HIRO": {
"LOWER_LEVEL": {
"max_lower_level_timesteps": 5,
"Actor": {
"learning_rate": 0.001,
"linear_hidden_units": [20, 20],
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"linear_hidden_units": [20, 20],
"final_layer_activation": "None",
"batch_norm": False,
"buffer_size": 100000,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"batch_size": 256,
"discount_rate": 0.9,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 20,
"learning_updates_per_learning_session": 10,
"clip_rewards": False
} ,
"HIGHER_LEVEL": {
"Actor": {
"learning_rate": 0.001,
"linear_hidden_units": [20, 20],
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"linear_hidden_units": [20, 20],
"final_layer_activation": "None",
"batch_norm": False,
"buffer_size": 100000,
"tau": 0.005,
"gradient_clipping_norm": 5
},
"batch_size": 256,
"discount_rate": 0.9,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 20,
"learning_updates_per_learning_session": 10,
"clip_rewards": False
} ,
},
"Actor_Critic_Agents": { # hyperparameters taken from https://arxiv.org/pdf/1802.09477.pdf
"Actor": {
"learning_rate": 0.001,
"linear_hidden_units": [400, 300],
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.01,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"linear_hidden_units": [400, 300],
"final_layer_activation": "None",
"batch_norm": False,
"buffer_size": 100000,
"tau": 0.01,
"gradient_clipping_norm": 5
},
"batch_size": 64,
"discount_rate": 0.99,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.2, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 1,
"learning_updates_per_learning_session": 1,
"clip_rewards": False
}
}
config.hyperparameters = {
"Policy_Gradient_Agents": {
"learning_rate": 0.05,
"linear_hidden_units": [30, 15],
"final_layer_activation": "TANH",
"learning_iterations_per_round": 10,
"discount_rate": 0.9,
"batch_norm": False,
"clip_epsilon": 0.2,
"episodes_per_learning_round": 10,
"normalise_rewards": True,
"gradient_clipping_norm": 5,
"mu": 0.0,
"theta": 0.15,
"sigma": 0.2,
"epsilon_decay_rate_denominator": 1,
"clip_rewards": clip_rewards
},
"Actor_Critic_Agents": actor_critic_agent_hyperparameters,
"DIAYN": {
"DISCRIMINATOR": {
"learning_rate": 0.001,
"linear_hidden_units": [32, 32],
"final_layer_activation": None,
"gradient_clipping_norm": 5
},
"AGENT": actor_critic_agent_hyperparameters,
"MANAGER": manager_hyperparameters,
"num_skills": 10,
"num_unsupservised_episodes": 100
}
}
config.hyperparameters = {
"Actor_Critic_Agents": {
"Actor": {
"learning_rate": 0.001,
"linear_hidden_units": [50, 50],
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.01,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"linear_hidden_units": [50, 50, 50],
"final_layer_activation": None,
"batch_norm": False,
"buffer_size": 30000,
"tau": 0.01,
"gradient_clipping_norm": 5
},
"batch_size": 256,
"discount_rate": 0.9,
"mu": 0.0,
"theta": 0.15,
"sigma": 0.25,
"update_every_n_steps": 10,
"learning_updates_per_learning_session": 10,
"HER_sample_proportion": 0.8,
"clip_rewards": False
}
}
config.runs_per_agent = 3
config.use_GPU = False
config.overwrite_existing_results_file = False
config.randomise_random_seed = True
config.save_model = False
config.hyperparameters = {
"DQN_Agents": {
"learning_rate": 0.001,
"batch_size": 128,
"buffer_size": 100000,
"epsilon_decay_rate_denominator": 150,
"discount_rate": 0.999,
"incremental_td_error": 1e-8,
"update_every_n_steps": 1,
"linear_hidden_units": [64, 64],
"final_layer_activation": None,
"y_range": (-1, 14),
"batch_norm": False,
"gradient_clipping_norm": 5,
"HER_sample_proportion": 0.8,
"learning_iterations": 1,
"clip_rewards": False
}
}
if __name__ == '__main__':
AGENTS = [DQN_HER, DQN]
trainer = Trainer(config, AGENTS)
trainer.run_games_for_agents()
config.hyperparameters = {
"h_DQN": {
"CONTROLLER": {
"batch_size": 256,
"learning_rate": 0.01,
"buffer_size": 40000,
"linear_hidden_units": [20, 10],
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0, 1],
"embedding_dimensions": [[config.environment.observation_space.n,
max(4, int(config.environment.observation_space.n / 10.0))],
[config.environment.observation_space.n,
max(4, int(config.environment.observation_space.n / 10.0))]],
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 1500,
"discount_rate": 0.999,
"learning_iterations": 1
},
"META_CONTROLLER": {
"batch_size": 256,
"learning_rate": 0.001,
"buffer_size": 40000,
"linear_hidden_units": [20, 10],
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": [[config.environment.observation_space.n,
max(4, int(config.environment.observation_space.n / 10.0))]],
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 2500,
"discount_rate": 0.999,
"learning_iterations": 1
}
},
"SNN_HRL": {
"SKILL_AGENT": {
"num_skills": 2,
"regularisation_weight": 1.5,
"visitations_decay": 0.99,
"episodes_for_pretraining": 2000,
# "batch_size": 256,
# "learning_rate": 0.01,
# "buffer_size": 40000,
# "linear_hidden_units": [20, 10],
# "final_layer_activation": "None",
# "columns_of_data_to_be_embedded": [0, 1],
# "embedding_dimensions": [[config.environment.observation_space.n,
# max(4, int(config.environment.observation_space.n / 10.0))],
# [6, 4]],
# "batch_norm": False,
# "gradient_clipping_norm": 5,
# "update_every_n_steps": 1,
# "epsilon_decay_rate_denominator": 50,
# "discount_rate": 0.999,
# "learning_iterations": 1
"learning_rate": 0.05,
"linear_hidden_units": [20, 20],
"final_layer_activation": "SOFTMAX",
"learning_iterations_per_round": 5,
"discount_rate": 0.99,
"batch_norm": False,
"clip_epsilon": 0.1,
"episodes_per_learning_round": 4,
"normalise_rewards": True,
"gradient_clipping_norm": 7.0,
"mu": 0.0, # only required for continuous action games
"theta": 0.0, # only required for continuous action games
"sigma": 0.0, # only required for continuous action games
"epsilon_decay_rate_denominator": 1.0
},
"MANAGER": {
"timesteps_before_changing_skill": 4,
"linear_hidden_units": [10, 5],
"learning_rate": 0.01,
"buffer_size": 40000,
"batch_size": 256,
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": [[config.environment.observation_space.n,
max(4, int(config.environment.observation_space.n / 10.0))]],
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 1000,
"discount_rate": 0.999,
"learning_iterations": 1
}
}
}
config.hyperparameters = {
"DQN_Agents": {
"linear_hidden_units": [30, 10],
"learning_rate": 0.01,
"buffer_size": 40000,
"batch_size": 256,
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": embedding_dimensions,
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 10,
"discount_rate": 0.99,
"learning_iterations": 1,
"tau": 0.01,
"exploration_cycle_episodes_length": None,
"learning_iterations": 1,
"clip_rewards": False
},
"SNN_HRL": {
"SKILL_AGENT": {
"num_skills": 20,
"regularisation_weight": 1.5,
"visitations_decay": 0.9999,
"episodes_for_pretraining": 300,
"batch_size": 256,
"learning_rate": 0.001,
"buffer_size": 40000,
"linear_hidden_units": [20, 10],
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0, 1],
"embedding_dimensions": [embedding_dimensions[0], [20, 6]],
"batch_norm": False,
"gradient_clipping_norm": 2,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 500,
"discount_rate": 0.999,
"learning_iterations": 1,
"tau": 0.01,
"clip_rewards": False
},
"MANAGER": {
"timesteps_before_changing_skill": 6,
"linear_hidden_units": [10, 5],
"learning_rate": 0.01,
"buffer_size": 40000,
"batch_size": 256,
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": embedding_dimensions,
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 50,
"discount_rate": 0.99,
"learning_iterations": 1,
"tau": 0.01,
"clip_rewards": False
}
},
"Actor_Critic_Agents": {
"learning_rate": 0.005,
"linear_hidden_units": [20, 10],
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": embedding_dimensions,
"final_layer_activation": ["SOFTMAX", None],
"gradient_clipping_norm": 5.0,
"discount_rate": 0.99,
"epsilon_decay_rate_denominator": 50.0,
"normalise_rewards": True,
"clip_rewards": False
},
"DIAYN": {
"num_skills": 5,
"DISCRIMINATOR": {
"learning_rate": 0.01,
"linear_hidden_units": [20, 10],
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": embedding_dimensions,
},
"AGENT": {
"learning_rate": 0.01,
"linear_hidden_units": [20, 10],
}
},
"HRL": {
"linear_hidden_units": [10, 5],
"learning_rate": 0.01,
"buffer_size": 40000,
"batch_size": 256,
"final_layer_activation": "None",
"columns_of_data_to_be_embedded": [0],
"embedding_dimensions": embedding_dimensions,
"batch_norm": False,
"gradient_clipping_norm": 5,
"update_every_n_steps": 1,
"epsilon_decay_rate_denominator": 400,
"discount_rate": 0.99,
"learning_iterations": 1,
"tau": 0.01
}
}
config.hyperparameters = {
"Policy_Gradient_Agents": {
"learning_rate": 0.05,
"linear_hidden_units": [30, 15],
"final_layer_activation": "TANH",
"learning_iterations_per_round": 10,
"discount_rate": 0.9,
"batch_norm": False,
"clip_epsilon": 0.2,
"episodes_per_learning_round": 10,
"normalise_rewards": True,
"gradient_clipping_norm": 5,
"mu": 0.0,
"theta": 0.15,
"sigma": 0.2,
"epsilon_decay_rate_denominator": 1,
"clip_rewards": False
},
"Actor_Critic_Agents": {
"Actor": {
"learning_rate": 0.003,
"linear_hidden_units": [20, 20],
"final_layer_activation": None,
"batch_norm": False,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"Critic": {
"learning_rate": 0.02,
"linear_hidden_units": [20, 20],
"final_layer_activation": None,
"batch_norm": False,
"buffer_size": 1000000,
"tau": 0.005,
"gradient_clipping_norm": 5,
"initialiser": "Xavier"
},
"min_steps_before_learning": 1000, #for SAC only
"batch_size": 256,
"discount_rate": 0.99,
"mu": 0.0, # for O-H noise
"theta": 0.15, # for O-H noise
"sigma": 0.25, # for O-H noise
"action_noise_std": 0.2, # for TD3
"action_noise_clipping_range": 0.5, # for TD3
"update_every_n_steps": 20,
"learning_updates_per_learning_session": 10,
"automatically_tune_entropy_hyperparameter": True,
"entropy_term_weight": None,
"add_extra_noise": True,
"do_evaluation_iterations": True,
"clip_rewards": False
}
}