def compute(self, config, budget, working_directory, *args, **kwargs):
"""
Simple example for a compute function using a feed forward network.
It is trained on the MNIST dataset.
The input parameter "config" (dictionary) contains the sampled
configurations passed by the bohb optimizer
"""
env = ContinuousCartPoleEnv(reward_function=smooth_reward)
state_dim = env.observation_space.shape[0]
# Try to ensure determinism
############################
torch.manual_seed(config['seed'])
env.seed(config['seed'])
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
############################
# conf dictionary to controll training
conf = {'lr':config['lr'], 'bs':64, 'loss':nn.MSELoss(),
'hidden_dim':config['hidden_dim'],
'mem_size':50000, 'activation':config['activation'],
'epsilon':config['epsilon'],
'eps_scheduler':'exp', 'n_episodes':budget,
'dropout_rate': config['dropout_rate'], 'n_cycles': 1,
'decay_rate': config['decay_rate']
}
############################
# create dqn object and train it
dqn = DQN(state_dim, config['action_dim'],
gamma=config['gamma'], conf=conf)
time_steps = 1000
stats = dqn.train(int(budget), time_steps, env, conf)
# plot_episode_stats(stats, noshow=True)
final_reward = 0
for _ in range(5):
s = env.reset()
for _ in range(time_steps):
# env.render()
action = dqn.get_action(s, 0.)
s, r, d, _ = env.step(dqn.action.act(action)) #(action - 5)/6]))
final_reward += r
if d:
break
env.close()
# ###########################
return ({
# remember: HpBandSter always minimizes!
'loss': - (final_reward / 5),
'info': {'max_len_train': max(stats.episode_lengths),
'max_reward_train': max(stats.episode_rewards),
'avg_final': (final_reward / 5) }
})
# When is done, executes
if done:
print("Episode finished after {} timesteps".format(t + 1))
rewards_list.append(total_reward)
steps_list.append(t)
break
print("Total reward in this episode: ", total_reward)
print("")
print("========== RESULTS ==========")
print("=============================")
print("Total number of episodes: ", num_episodes)
print("Average reward of all episodes: ", np.average(rewards_list))
print("")
env.close()
############################
# Plots
############################
'''
# summarize history for loss
fig1 = plt.figure()
plt.plot(loss_list)
plt.title('Model Loss')
plt.ylabel('loss')
plt.xlabel('steps')
#plt.show()
fig1.savefig('model_loss.png')
'''