def _build_for_dueling_dqn(self) -> Dict[str, Any]:
config_dict = self._build_for_dqn()
config_dict.update({'name': os.path.join(self.folder, 'DuelingDQN'),
'model_architecture': ConvNN(observation_shape=(84, 84, self.frame_depth), dueling=True,
n_actions=6, opt='adam', learning_rate=0.000102)})
return config_dict
def _build_for_dqn(self) -> Dict[str, Any]:
return {
'name':
os.path.join(self.folder, 'DeepQAgent'),
'env_spec':
self.env_spec,
'env_wrappers':
self.env_wrappers,
'model_architecture':
ConvNN(observation_shape=(84, 84, self.frame_depth),
n_actions=6,
output_activation=None,
opt='adam',
learning_rate=0.00008),
'gamma':
0.99,
'final_reward':
None,
'eps':
EpsilonGreedy(eps_initial=2,
decay=0.000025,
eps_min=0.01,
decay_schedule='linear'),
'replay_buffer':
ContinuousBuffer(buffer_size=40000),
'replay_buffer_samples':
32
}
def _build_for_dqn(self) -> Dict[str, Any]:
return {
'name':
os.path.join(self.folder, 'DeepQAgent'),
'env_spec':
self.env_spec,
'env_wrappers':
self.env_wrappers,
'model_architecture':
ConvNN(observation_shape=(84, 84, self.frame_depth),
n_actions=6,
output_activation=None,
opt='adam',
learning_rate=0.000105),
'gamma':
0.99,
'final_reward':
None,
# Use eps_initial > 1 here so only random actions used for first steps, which will make filling the
# replay buffer more efficient. It'll also avoid decaying eps while not training.
# Alternative: 'eps': EpsilonGreedy(eps_initial=1.2, decay=0.000025, eps_min=0.01,
# decay_schedule='compound'),
'eps':
EpsilonGreedy(eps_initial=1.1,
decay=0.00001,
eps_min=0.01,
decay_schedule='linear'),
'replay_buffer':
ContinuousBuffer(buffer_size=10000),
'replay_buffer_samples':
32
}
def _build_for_ac(self) -> Dict[str, Any]:
return {'name': os.path.join(self.folder, 'ActorCriticAgent'),
'env_spec': self.env_spec,
'env_wrappers': self.env_wrappers,
'model_architecture': ConvNN(observation_shape=(84, 84, self.frame_depth), n_actions=6,
output_activation='softmax', opt='adam', learning_rate=0.000105,
output_type='ac'),
'gamma': 0.99,
'final_reward': None}
def _build_for_double_dueling_dqn(self) -> Dict[str, Any]:
config_dict = self._build_for_dqn()
config_dict.update({'name': os.path.join(self.folder, 'DoubleDuelingDQN'),
'double': True,
'model_architecture': ConvNN(
observation_shape=(self.target_obs_shape[0], self.target_obs_shape[1],
self.frame_depth), dueling=True,
n_actions=self.n_actions, opt='adam', learning_rate=0.0001)})
return config_dict