def test_default_algo(use_ref_model: bool = False):
env = utils_env.Environment()
# use default params
ag = agent.DRLAgent()
if use_ref_model:
print('... Test the agent using reference model ...')
ag.set_model_path('ref')
al = algo.DRLAlgo(env, ag)
al.test()
def fit(self, i: int, env: utils_env.Environment):
self.ag = agent.DRLAgent(state_size, action_size,
self.memory_size, self.gamma, self.batch_size,
self.expl_noise, self.model_learning_rate,
self.num_fc_actor, self.num_fc_critic)
self.ag.set_model_path(i) # save each candidate's model separately
self.al = algo.DRLAlgo(env, self.ag, self.num_episodes)
self.al.set_image_path(i) # save each candidate's score separately
history = self.al.train(with_close=False) # do not close the Env so that other agents can be trained
score = self._get_score(history)
return score
def test_default_algo(use_ref_model: bool = False):
env = utils_env.Environment()
model_name_suffix = ''
if use_ref_model:
print('... Test the agent using reference model ...')
model_name_suffix = 'ref_'
# use default params
ag_1 = agent.DRLAgent()
ag_1.set_model_path(model_name_suffix + str(1))
ag_2 = agent.DRLAgent()
ag_2.set_model_path(model_name_suffix + str(2))
al = algo.DRLAlgo(env, ag_1, ag_2)
al.test()
def train_two_agents():
env = utils_env.Environment()
# use default params
ag_1 = agent.DRLAgent()
ag_1.set_model_path(1)
ag_2 = agent.DRLAgent()
ag_2.set_model_path(2)
al = algo.DRLAlgo(env, ag_1, ag_2)
history, best_e, best_score = al.train()
print('\nFinal score: {:.3f}'.format(
np.mean(history[-const.rolling_mean_N:])))
print('Final memory length:', ag_1.memory.get_length())
print('Best score in {:d} episodes, avg_score: {:.3f}'.format(
best_e, best_score))
# plot losses
losses_lists = [
ag_1.actor_loss_list, ag_2.actor_loss_list, ag_1.critic_loss_list,
ag_2.critic_loss_list
]
losses_labels = [
'agent_1_actor', 'agent_2_actor', 'agent_1_critic', 'agent_2_critic'
]
utils_plot.plot_loss(losses_lists, losses_labels)
# plot noise
utils_plot.plot_scatter(ag_1.noise_list,
title_text='Noise',
fp=const.file_path_img_noise)
# plot memory actions
memory_actions = np.array([t[1] for t in ag_1.memory.memory])
utils_plot.plot_scatter(memory_actions,
title_text='Actions',
fp=const.file_path_img_actions)
# show mean memory actions
mean_a = np.mean(memory_actions, axis=0)
std_a = np.std(memory_actions, axis=0)
print('Mean/std actions agent_1:', mean_a[:2], std_a[:2])
print('Mean/std actions agent_2:', mean_a[2:], std_a[2:])
def fit(self, i: int, env: utils_env.Environment):
self.ag_1 = agent.DRLAgent(self.num_states, self.num_actions,
self.memory_size, self.gamma,
self.batch_size, self.tau,
self.model_learning_rate, self.num_fc_1,
self.num_fc_2)
self.ag_1.set_model_path(
str(i) + '_0') # save each candidate's model separately
self.ag_2 = agent.DRLAgent(self.num_states, self.num_actions,
self.memory_size, self.gamma,
self.batch_size, self.tau,
self.model_learning_rate, self.num_fc_1,
self.num_fc_2)
self.ag_2.set_model_path(
str(i) + '_1') # save each candidate's model separately
self.al = algo.DRLAlgo(env, self.ag_1, self.ag_2, self.num_episodes)
self.al.set_image_path(i) # save each candidate's score separately
# do not close the Env so that other agents can be trained
history, _, best_score = self.al.train(
with_close=False) # todo use best score?
score = self._get_score(history)
return score
def train_default_algo():
env = utils_env.Environment()
# use default params
ag = agent.DRLAgent()
al = algo.DRLAlgo(env, ag)
al.train()