def collect_stats(agent: DQNAgent, n_games=1000):
MAX_STEPS = 1000
lenghts = []
looped = 0
for i in range(1, n_games+1):
env = gym.make('snake-v0')
# env.__init__(human_mode=False)
observation = env.reset()
done = False
steps = 0
agent.epsilon = 0.0
state = agent.get_last_observations(observation)
while not done and steps < MAX_STEPS:
action = agent.act(state)
next_observation, _, done, _ = env.step(action)
state = agent.get_last_observations(next_observation)
steps += 1
if steps == MAX_STEPS:
looped += 1
else:
lenghts.append(len(env.game.snake.body))
if i % (n_games//10) == 0:
print(f"Avg len: {sum(lenghts) / len(lenghts):.2f}, looped {looped}/{i}")
def watch_agent(agent: DQNAgent):
env = gym.make('snake-v0')
env.__init__(human_mode=True)
observation = env.reset()
renderer=Renderer(env.game)
try:
done = False
steps = 0
agent.epsilon = 0
state = agent.get_last_observations(observation)
while not done:
# time.sleep(0.001)
renderer.render_frame()
action = agent.act(state)
next_observation, _, done, _ = env.step(action)
state = agent.get_last_observations(next_observation)
steps += 1
finally:
renderer.close_window()
print(f"Snake length: {len(env.game.snake.body)}")
print(f"Simulation ended after {steps} steps.")
target_network = DQNAgent(state_shape, n_actions).to(device)
target_network.load_state_dict(agent.state_dict())
opt = torch.optim.Adam(agent.parameters(), lr=1e-4)
exp_replay = ReplayBuffer(buffer_size)
print('test_buffer')
for i in range(100):
play_and_record(state, agent, env, exp_replay, n_steps=10**2)
if len(exp_replay) == buffer_size:
break
print(len(exp_replay))
state = env.reset()
for step in trange(step, total_steps + 1):
agent.epsilon = linear_decay(init_epsilon, final_epsilon, step,
decay_steps)
# play
_, state = play_and_record(state, agent, env, exp_replay,
timesteps_per_epoch)
# train
obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(
batch_size)
loss = compute_td_loss(obs_batch,
act_batch,
reward_batch,
next_obs_batch,
is_done_batch,
agent,
