def make_agent(process_idx):
n_hidden_channels = 50
if self.use_lstm:
q_func = FCLSTMStateQFunction(
ndim_obs,
n_actions,
n_hidden_channels=n_hidden_channels,
n_hidden_layers=2)
else:
q_func = FCStateQFunctionWithDiscreteAction(
ndim_obs,
n_actions,
n_hidden_channels=n_hidden_channels,
n_hidden_layers=2)
opt = rmsprop_async.RMSpropAsync(lr=1e-3, eps=1e-2, alpha=0.99)
opt.setup(q_func)
if self.explorer == 'epsilon_greedy':
explorer = chainerrl.explorers.ConstantEpsilonGreedy(
process_idx / 10, random_action_func)
else:
explorer = chainerrl.explorers.Boltzmann()
return nsq.NSQ(q_func,
opt,
t_max=self.t_max,
gamma=0.9,
i_target=100,
explorer=explorer)
def make_agent(process_idx):
# Random epsilon assignment described in the original paper
rand = random.random()
if rand < 0.4:
epsilon_target = 0.1
elif rand < 0.7:
epsilon_target = 0.01
else:
epsilon_target = 0.5
explorer = explorers.LinearDecayEpsilonGreedy(
1, epsilon_target, args.final_exploration_frames,
action_space.sample)
# Suppress the explorer logger
explorer.logger.setLevel(logging.INFO)
return nsq.NSQ(q_func, opt, t_max=5, gamma=0.99,
i_target=40000,
explorer=explorer, phi=dqn_phi)