def update(self, action_a, reward_a, state_a, done_a):
'''
Update per timestep after env transitions, e.g. memory, algorithm, update agent params, train net
'''
for (e, b), body in util.ndenumerate_nonan(self.body_a):
body.memory.update(
action_a[(e, b)], reward_a[(e, b)], state_a[(e, b)], done_a[(e, b)])
loss_a = self.algorithm.train()
loss_a = util.guard_data_a(self, loss_a, 'loss')
explore_var_a = self.algorithm.update()
explore_var_a = util.guard_data_a(self, explore_var_a, 'explore_var')
return loss_a, explore_var_a
def update(self, action_a, reward_a, state_a, done_a):
'''
Update per timestep after env transitions, e.g. memory, algorithm, update agent params, train net
'''
for (e, b), body in util.ndenumerate_nonan(self.body_a):
body.memory.update(action_a[(e, b)], reward_a[(e, b)], state_a[(e, b)], done_a[(e, b)])
loss_a = self.algorithm.train()
loss_a = util.guard_data_a(self, loss_a, 'loss')
for (e, b), body in util.ndenumerate_nonan(self.body_a):
body.loss = loss_a[(e, b)]
explore_var_a = self.algorithm.update()
explore_var_a = util.guard_data_a(self, explore_var_a, 'explore_var')
return loss_a, explore_var_a
def space_update(self, action_a, reward_a, state_a, done_a):
'''Update per timestep after env transitions, e.g. memory, algorithm, update agent params, train net'''
for eb, body in util.ndenumerate_nonan(self.body_a):
body.action_pd_update()
body.memory.update(action_a[eb], reward_a[eb], state_a[eb],
done_a[eb])
loss_a = self.algorithm.space_train()
loss_a = util.guard_data_a(self, loss_a, 'loss')
for eb, body in util.ndenumerate_nonan(self.body_a):
if not np.isnan(loss_a[eb]): # set for log_summary()
body.loss = loss_a[eb]
explore_var_a = self.algorithm.space_update()
explore_var_a = util.guard_data_a(self, explore_var_a, 'explore_var')
logger.debug(
f'Agent {self.a} loss: {loss_a}, explore_var_a {explore_var_a}')
for eb, body in util.ndenumerate_nonan(self.body_a):
if body.env.done:
body.epi_update()
return loss_a, explore_var_a