# set new weights
agent.set_weights(new_weights)
manager.set_agent(new_weights)
# get new weights
agent = manager.get_agent()
# update aggregator
time_steps = manager.test(test_steps)
manager.update_aggregator(loss=loss, time_steps=time_steps)
# print progress
print(
f"epoch ::: {e} loss ::: {loss} avg env steps ::: {np.mean(time_steps)}"
)
# you can also alter your managers parameters
if e % 5 == 0:
epsilon = epsilon * .9
manager.set_epsilon(epsilon=epsilon)
print(f"New epsilon: {epsilon}")
# if e % saving_after == 0:
# #you can save models
# manager.save_model(saving_path, e)
# and load mmodels
# manager.load_model(saving_path)
print("done")
print("testing optimized agent")
manager.test(test_steps, test_episodes=10, render=True, do_print=True)
# get and apply new weights
new_weights = agent.get_weights()
manager.set_agent(new_weights)
# get new agent
agent = manager.get_agent()
# update aggregator
time_steps = manager.test(test_steps,
render=True if e % 5 == 0 else False)
manager.update_aggregator(loss=losses_list, time_steps=time_steps)
print(
f"epoch ::: {e} loss ::: {np.mean([np.mean(l) for l in losses_list])} avg env steps ::: {np.mean(time_steps)}"
)
manager.agg.save_graphic()
print("---")
# alter epsilon parameter
new_epsilon = 0.90 + (0.05 / (e + 1))
manager.set_epsilon(epsilon=new_epsilon)
# test after training
print("done")
print("testing optimized agent")
manager.test(test_steps,
test_episodes=10,
render=True,
do_print=True,
evaluation_measure="time_and_reward")
np.mean(np.random.normal(size=(64, 100)), axis=0)
for _ in range(1000)
]
new_weights = agent.model.get_weights()
# set new weights
manager.set_agent(new_weights)
# get new weights
agent = manager.get_agent()
# update aggregator
time_steps = manager.test(test_steps)
manager.update_aggregator(loss=dummy_losses, time_steps=time_steps)
# print progress
print(
f"epoch ::: {e} loss ::: {np.mean([np.mean(l) for l in dummy_losses])} avg env steps ::: {np.mean(time_steps)}"
)
# yeu can also alter your managers parameters
manager.set_epsilon(epsilon=0.99)
if e % saving_after == 0:
# you can save models
manager.save_model(saving_path, e)
# and load mmodels
manager.load_model(saving_path)
print("done")
print("testing optimized agent")
manager.test(test_steps, test_episodes=10, render=True)
# get new weights
agent = manager.get_agent()
# update aggregator
time_steps = manager.test(test_steps, test_episodes=50, render=False)
manager.update_aggregator(loss=loss, time_steps=time_steps)
print(
f"epoch ::: {e} loss ::: {loss.numpy()} avg env steps ::: {np.mean(time_steps)}"
)
# store performance
mean_list.append(np.mean(time_steps))
# if performace was good 5 times in a row, stop training
if np.mean(mean_list) > 195:
break
# Annealing epsilon
if (e + 1) % 5 == 0:
new_epsilon = 0.85 * manager.kwargs['epsilon']
manager.set_epsilon(new_epsilon)
print("New Epsilon: ", new_epsilon)
print("Done!")
print("Testing optimized agent...")
manager.set_epsilon(0.0)
manager.test(test_steps, test_episodes=10, render=True, do_print=True)