config.hyperparameters = {
"Policy_Gradient_Agents": {
"learning_rate": 0.02,
"nn_layers": 2,
"nn_start_units": 20,
"nn_unit_decay": 1.0,
"final_layer_activation": "TANH",
"learning_iterations_per_round": 10,
"discount_rate": 0.99,
"batch_norm": False,
"clip_epsilon": 0.2,
"episodes_per_learning_round": 7,
"normalise_rewards": True,
"gradient_clipping_norm": 5,
"mu": 0.0,
"theta": 0.15,
"sigma": 0.25,
"noise_decay_denominator": 1
},
"Actor_Critic_Agents": {
"Actor": {
"learning_rate": 0.001,
"nn_layers": 2,
"nn_start_units": 20,
"nn_unit_decay": 1.0,
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.001,
"gradient_clipping_norm": 1
},
"Critic": {
"learning_rate": 0.01,
"nn_layers": 2,
"nn_start_units": 20,
"nn_unit_decay": 1.0,
"final_layer_activation": None,
"batch_norm": False,
"buffer_size": 100000,
"tau": 0.001,
"gradient_clipping_norm": 1
},
"batch_size": 256,
"discount_rate": 0.99,
"mu": 0.0,
"theta": 0.15,
"sigma": 0.25, #0.22 did well before
"noise_decay_denominator": 5,
"update_every_n_steps": 10,
"learning_updates_per_learning_session": 5
}
}
config.hyperparameters = {
"Actor_Critic_Agents": {
"Actor": {
"learning_rate": 0.001,
"nn_layers": 5,
"nn_start_units": 50,
"nn_unit_decay": 1.0,
"final_layer_activation": "TANH",
"batch_norm": False,
"tau": 0.01,
"gradient_clipping_norm": 5
},
"Critic": {
"learning_rate": 0.01,
"nn_layers": 6,
"nn_start_units": 50,
"nn_unit_decay": 1.0,
"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, # 0.22 did well before
"update_every_n_steps": 10,
"learning_updates_per_learning_session": 10
}
}
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
}
}
config.hyperparameters = {
"DQN_Agents": {
"learning_rate": 0.005,
"batch_size": 256,
"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": 1,
"nn_layers": 3,
"nn_start_units": 20,
"nn_unit_decay": 1.0,
"final_layer_activation": None,
"batch_norm": False,
"gradient_clipping_norm": 5
},
"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
},
"Policy_Gradient_Agents": {
"learning_rate": 0.05,
"nn_layers": 2,
"nn_start_units": 20,
"nn_unit_decay": 1.0,
"final_layer_activation": "SOFTMAX",
"learning_iterations_per_round": 10,
"discount_rate": 0.99,
"batch_norm": False,
"clip_epsilon": 0.1,
"episodes_per_learning_round": 7,
"normalise_rewards": True,
"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
"noise_decay_denominator": 1 #only required for continuous action games
}
}