def __init__(self, env_name):
self.device = torch.device('cuda')
self.env_name = env_name
self.env = wrappers.make_env(env_name)
self.gamma = 0.99
self.batch_size = 32
self.replay_buffer_size = 10000
self.replay_start_size = 10000
self.learning_rate = 1e-4
self.update_target_interval = 1000
self.epsilon_start = 1.0
self.epsilon_end = 0.02
self.epsilon_period = 100000
self.reward_bound = 19.5
self.replay_buffer = replay_buffer.ReplayBuffer(
self.replay_buffer_size)
self.network = dqn_model.DQNModel(self.env.observation_space.shape,
self.env.action_space.n).to(
self.device)
self.target_network = dqn_model.DQNModel(
self.env.observation_space.shape,
self.env.action_space.n).to(self.device)
self.optimizer = optim.Adam(self.network.parameters(),
lr=self.learning_rate)
print(self.network)
self.writer = SummaryWriter(comment='dqn' + self.env_name)
self.total_rewards = []
self.frame_index = 0
def __init__(self, env_name):
self.env = wrappers.make_env(env_name)
self.env_name = env_name
self.device = torch.device('cuda')
self.learning_rate = 2.5e-4
self.stabilizer = 0.01
self.gradient_momentum = 0.95
self.gamma = 0.99
self.batch_size = 32
self.replay_start_size = 50000
self.replay_buffer_size = 1000000
self.update_target_interval = 10000
self.training_frequency = 4
self.epsilon_start = 1.0
self.epsilon_end = 0.05
self.epsilon_period = 1000000
self.network = dqn_model.DQNModel(self.env.observation_space.shape, self.env.action_space.n).to(self.device)
self.target_network = dqn_model.DQNModel(self.env.observation_space.shape, self.env.action_space.n).to(self.device)
self.replay_buffer = replay_buffer.ReplayBuffer(self.replay_buffer_size)
self.optimizer = optim.RMSprop(self.network.parameters(), lr = self.learning_rate, alpha = self.gradient_momentum,
eps = self.stabilizer)
self.loss_criterion = nn.SmoothL1Loss()
print(self.network)
self.writer = SummaryWriter(comment = 'dqn' + self.env_name)
self.total_rewards = []
self.best_mean_reward = None
self.steps = 0
def __init__(self, env_name):
self.env = wrappers.make_env(env_name)
self.env_name = env_name
self.device = torch.device('cuda')
torch.manual_seed(2)
np.random.seed(2)
self.env.seed(2)
self.env.action_space.seed(2)
self.env.observation_space.seed(2)
self.learning_rate = 1e-4
self.gamma = 0.99
self.batch_size = 32
self.replay_start_size = 10000
self.replay_buffer_size = 10000
self.update_target_interval = 1000
self.epsilon_start = 1.0
self.epsilon_end = 0.02
self.epsilon_period = 100000
self.alpha = 0.6
self.beta_start = 0.4
self.beta_period = 100000
self.beta = self.beta_start
self.network = dqn_model.DQNModel(self.env.observation_space.shape,
self.env.action_space.n).to(
self.device)
self.target_network = dqn_model.DQNModel(
self.env.observation_space.shape,
self.env.action_space.n).to(self.device)
self.replay_buffer = replay_buffer.PriorityBuffer(
self.replay_buffer_size, self.alpha)
self.optimizer = optim.Adam(self.network.parameters(),
lr=self.learning_rate)
print(self.network)
self.writer = SummaryWriter(comment='dqnpriority' + self.env_name)
self.total_rewards = []
self.best_mean_reward = None
self.steps = 0