def test_agent_solve_bit_flipping_game():
AGENTS = [PPO, DDQN, DQN_With_Fixed_Q_Targets, DDQN_With_Prioritised_Experience_Replay, DQN, DQN_HER]
trainer = Trainer(config, AGENTS)
results = trainer.run_games_for_agents()
for agent in AGENTS:
agent_results = results[agent.agent_name]
agent_results = np.max(agent_results[0][1][50:])
assert agent_results >= 0.0, "Failed for {} -- score {}".format(agent.agent_name, agent_results)
def test_agents_can_play_games_of_different_dimensions():
config.num_episodes_to_run = 10
config.hyperparameters["DQN_Agents"]["batch_size"] = 3
AGENTS = [A2C, A3C, PPO, DDQN, DQN_With_Fixed_Q_Targets, DDQN_With_Prioritised_Experience_Replay, DQN]
trainer = Trainer(config, AGENTS)
config.environment = gym.make("CartPole-v0")
results = trainer.run_games_for_agents()
for agent in AGENTS:
assert agent.agent_name in results.keys()
AGENTS = [SAC, TD3, PPO, DDPG]
config.environment = gym.make("MountainCarContinuous-v0")
trainer = Trainer(config, AGENTS)
results = trainer.run_games_for_agents()
for agent in AGENTS:
assert agent.agent_name in results.keys()
AGENTS = [DDQN, SNN_HRL]
config.environment = Four_Rooms_Environment(15, 15, stochastic_actions_probability=0.25,
random_start_user_place=True, random_goal_place=False)
trainer = Trainer(config, AGENTS)
results = trainer.run_games_for_agents()
for agent in AGENTS:
assert agent.agent_name in results.keys()
def test_get_mean_and_standard_deviation_difference_results():
"""Tests that get_mean_and_standard_deviation_difference_results method produces correct output"""
results = [[1.0, 2.0, 3.0], [5.0, -33.0, 55.0], [2.5, 2.5, 2.5]]
mean_results = [
np.mean([1.0, 5.0, 2.5]),
np.mean([2.0, -33.0, 2.5]),
np.mean([3.0, 55.0, 2.5])
]
std_results = [
np.std([1.0, 5.0, 2.5]),
np.std([2.0, -33.0, 2.5]),
np.std([3.0, 55.0, 2.5])
]
mean_minus_1_std = [
mean - std_val for mean, std_val in zip(mean_results, std_results)
]
mean_plus_1_std = [
mean + std_val for mean, std_val in zip(mean_results, std_results)
]
config = Config()
config.standard_deviation_results = 1.0
trainer = Trainer(config, [])
mean_minus_x_std_guess, mean_results_guess, mean_plus_x_std_guess = trainer.get_mean_and_standard_deviation_difference_results(
results)
assert mean_results == mean_results_guess
assert mean_minus_1_std == mean_minus_x_std_guess
assert mean_plus_1_std == mean_plus_x_std_guess
config.standard_deviation_results = 3.0
trainer = Trainer(config, [])
mean_minus_x_std_guess, mean_results_guess, mean_plus_x_std_guess = trainer.get_mean_and_standard_deviation_difference_results(
results)
mean_plus_3_std = [
mean + 3.0 * std_val
for mean, std_val in zip(mean_results, std_results)
]
mean_minus_3_std = [
mean - 3.0 * std_val
for mean, std_val in zip(mean_results, std_results)
]
assert mean_results == mean_results_guess
assert mean_minus_3_std == mean_minus_x_std_guess
assert mean_plus_3_std == mean_plus_x_std_guess
"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
}
}
if __name__ == "__main__":
AGENTS = [
HRL
] #SAC_Discrete, SAC_Discrete, DDQN] #HRL] #, SNN_HRL, DQN, h_DQN]
trainer = Trainer(config, AGENTS)
trainer.run_games_for_agents()
"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"]
agent = DQN_HER(config)
agent.reset_game()
def test_initiation():
"""Tests whether DQN_HER initiates correctly"""
config.hyperparameters["batch_size"] = 64
agent = DQN_HER(config)
agent.reset_game()
assert agent.ordinary_buffer_batch_size == int(0.2 * 64)
assert agent.HER_buffer_batch_size == 64 - int(0.2 * 64)
assert agent.q_network_local.input_dim == 8