batch_size = 100
additional_episodes = math.ceil(batch_size / n_steps)
n_episodes += additional_episodes
# Loop over episodes
while episode_counter < n_episodes:
agent.reset()
loss_avg = 0
for step_num in range(n_steps):
transition = agent.step(dqn, epsilon)
replay_buffer.add(transition)
if replay_buffer.get_len() >= batch_size:
batch_transitions = replay_buffer.sample(batch_size)
loss = dqn.train_q_network(batch_transitions, gamma, do_target)
loss_avg += loss
if epsilon_decay and epsilon > epsilon_min and (
episode_counter - additional_episodes) >= 0:
epsilon = epsilon - epsilon_step
# if show_env: time.sleep(0.1)
# if epsilon_decay and epsilon > epsilon_min:
# epsilon = epsilon - epsilon_step
episode_counter += 1
print(f'Episode: {episode_counter}')
