batch_size=64,
layer1_size=256,
layer2_size=128,
n_actions=3)
#agent.load_models()
# np.random.seed(1)
score_history = []
for i in range(50):
obs = env.reset()
done = False
score = 0
while not done:
act = agent.choose_action(obs)
print(act)
new_state, reward, done, info = env.step(act)
agent.remember(obs, act, reward, new_state, int(done))
agent.learn()
score += reward
obs = new_state
#env.render()
score_history.append(score)
if i % 10 == 0:
agent.save_models()
env.render()
print('episode ', i, 'score %.2f' % score,
'trailing 25 games avg %.3f' % np.mean(score_history[-25:]))
def main(env, episodes=500, max_steps=500, eps_decay=.99,
actor_lr=10**-6, critic_lr=10**-3, gamma=.9,
base_nodes=64, batch_size=128,theta=.4, sigma=.25):
with tf.Session() as sess:
# Initialize environment and constants
input_dim = env.state_dim
output_dim = env.action_dim
action_high = env.action_high
action_low = env.action_low
# Create DDPG Agent
agent = Agent(input_dim, output_dim, action_high, action_low,
actor_lr=actor_lr, critic_lr=critic_lr, gamma=gamma,
base_nodes=base_nodes, eps_decay=eps_decay,
batch_size=batch_size,theta=theta, sigma=sigma,
sess=sess)
sess.run(tf.global_variables_initializer())
agent.actor.update_target_network()
agent.critic.update_target_network()
# Prepare for episodes
c_losses, rewards, actions, Qs, states = [np.array([]) for i in range(5)]
for e in tqdm(range(episodes)):
# Reset episode
state = env.reset()
state = np.reshape(state, (-1, len(state)))
agent.noise.reset()
done = False
step_count = 0
total_reward = 0
while not done and step_count < max_steps:
# Action
action = agent.act(state)
next_state, reward, done = env.step(action)
next_state = np.reshape(next_state, (-1, len(next_state)))
# Learn
c_loss = agent.learn(state, action, reward, done, next_state)
# Save results
c_losses = np.append(c_losses, c_loss)
actions = np.append(actions, action)
states = np.append(states, state[0])
Qs = np.append(Qs, agent.critic.predict(state, action))
# Loop
state = next_state
step_count += 1
total_reward += reward
# Reduce exploration
if agent.eps > agent.min_eps:
agent.eps *= agent.eps_decay
rewards = np.append(rewards, total_reward)
return rewards, c_losses, actions, Qs