# LOGGING
writer = SummaryWriter(comment="-" + params["run_name"] + "-noisy")
# NETWORK
net = dqn_noisy_net.Network(env.observation_space.shape,
env.action_space.n).to(device)
tgt_net = agents.TargetNetwork(net)
# AGENT
selector = actions.ArgmaxActionSelector()
agent = agents.DQNAgent(net, selector, device=device)
# RUNNER
exp_source = runner.RunnerSourceFirstLast(
env, agent,
gamma=params["gamma"]) # increase the number of steps for the runner
buffer = ExperienceReplayBuffer(exp_source,
buffer_size=params["replay_size"])
optimizer = optim.Adam(net.parameters(), lr=params["learning_rate"])
frame_idx = 0
# TRAIN
with logger.RewardTracker(writer, params["stop_reward"]) as reward_tracker:
while True:
frame_idx += 1
buffer.populate(1)
new_rewards = exp_source.pop_total_rewards()
if new_rewards:
env = gym.make(ENV_ID)
test_env = gym.make(ENV_ID)
act_net = ddpg_mlp.DDPGActor(env.observation_space.shape[0],
env.action_space.shape[0]).to(device)
crt_net = ddpg_mlp.DDPGCritic(env.observation_space.shape[0],
env.action_space.shape[0]).to(device)
print(act_net)
print(crt_net)
tgt_act_net = ptan.agent.TargetNet(act_net)
tgt_crt_net = ptan.agent.TargetNet(crt_net)
writer = SummaryWriter(comment="-ddpg_" + args.name)
agent = agents.AgentDDPG(act_net, device=device)
exp_source = runner.RunnerSourceFirstLast(env,
agent,
gamma=GAMMA,
steps_count=1)
buffer = memory.ExperienceReplayBuffer(exp_source, buffer_size=REPLAY_SIZE)
act_opt = optim.Adam(act_net.parameters(), lr=LEARNING_RATE)
crt_opt = optim.Adam(crt_net.parameters(), lr=LEARNING_RATE)
frame_idx = 0
best_reward = None
with logger.RewardTracker(act_net, writer, 200) as tracker:
with ptan.common.utils.TBMeanTracker(writer,
batch_size=10) as tb_tracker:
while True:
frame_idx += 1
buffer.populate(1)
rewards_steps = exp_source.pop_rewards_steps()
# NETWORK
net = dqn_mlp_net.Network(observation_space,
action_space,
hidden_layer_size=64).to(device)
tgt_net = agents.TargetNetwork(net)
# AGENT
selector = actions.EpsilonGreedyActionSelector(
epsilon=params["epsilon_start"])
epsilon_tracker = logger.EpsilonTracker(selector, params)
agent = agents.DQNAgent(net, selector, device=device)
# RUNNER
exp_source = runner.RunnerSourceFirstLast(env,
agent,
gamma=params["gamma"],
steps_count=1)
buffer = ExperienceReplayBuffer(exp_source,
buffer_size=params["replay_size"])
optimizer = optim.Adam(net.parameters(), lr=params["learning_rate"])
frame_idx = 0
done = False
# TRAIN
with logger.RewardTracker(writer, params["stop_reward"]) as reward_tracker:
while True:
frame_idx += 1
buffer.populate(1)
epsilon_tracker.frame(frame_idx)
# LOGGING
writer = SummaryWriter(comment="-" + params["run_name"] + "-reinforce")
# NETWORK
net = dqn_mlp_net.Network(observation_space,
action_space,
hidden_layer_size=64).to(device)
# AGENT
agent = agents.PolicyGradientAgent(net,
preprocessor=utils.float32_preprocessor,
apply_softmax=True)
# RUNNER
exp_source = runner.RunnerSourceFirstLast(env,
agent,
gamma=params["gamma"])
optimizer = optim.Adam(net.parameters(), lr=params["learning_rate"])
total_rewards = []
step_idx = 0
done_episodes = 0
batch_episodes = 0
batch_states, batch_actions, batch_qvals = [], [], []
cur_rewards = []
with logger.RewardTracker(writer, params["stop_reward"]) as reward_tracker:
for step_idx, exp in enumerate(exp_source):
batch_states.append(exp.state)
batch_actions.append(int(exp.action))