def __init__(self, n_in_elements, n_actions, explor_rate=0.0):
'''
Q値の範囲が報酬体系によって負の値をとる場合、F.reluは負の値をとれないので、学習に適さない。
活性化関数は、負の値も取ることが可能なものを選択する必要がある。
例えば、F.leaky_relu等。勾配消失問題を考えると、これが良い感じ。
n_size_twn_status: TWN自身の状態
num_ray_out: TWNセンサーをCNN層で処理した結果得られるデータ数のタプル (CNN層の出力要素数, CNN層の出力チャンネル数)
n_size_eb_status: TWNセンサーでとらえた、EBの情報
n_actions: 離散アクション空間
explor_rate=0.0: 探索行動比率（現時点で未使用）
'''
super().__init__()
with self.init_scope():
self.l4 = links.MLP(n_in_elements, int(n_in_elements*1.2), (n_in_elements*2, int(n_in_elements*1.8), int(n_in_elements*1.5)), nonlinearity=F.leaky_relu)
self.l5 = links.MLP(int(n_in_elements*1.2)+4, 4, (n_in_elements, int(n_in_elements*0.8), (n_in_elements*2)//3), nonlinearity=F.leaky_relu)
local_action_links_list = []
for i in range(n_actions):
action_links = links.MLP(4, 1, (n_in_elements//2,), nonlinearity=F.leaky_relu)
local_action_links_list.append(action_links)
self.action_chain_list = chainer.ChainList(*local_action_links_list)
self.explor_rate = explor_rate
self.debug_info = None
def __init__(self,
n_size_twn_status,
num_ray,
n_size_eb_status,
n_actions,
explor_rate=0.0):
self.n_size_twn_status = n_size_twn_status
self.num_ray = num_ray
self.n_size_eb_status = n_size_eb_status
self.num_history = 1
self.n_clasfy_ray = 32
super().__init__()
with self.init_scope():
self.ml1 = links.MLP(self.num_ray,
self.n_clasfy_ray,
((self.num_ray // 3) * 2,
self.num_ray // 2, self.num_ray // 3),
nonlinearity=F.leaky_relu)
aaa = n_size_twn_status + self.n_clasfy_ray + n_size_eb_status
self.l4 = L.Linear(aaa, aaa) #クラス分類用
self.ml5 = links.MLP(aaa,
n_actions, (aaa, aaa, aaa, aaa, aaa),
nonlinearity=nl)
self.explor_rate = explor_rate
self.debug_info = None
def __init__(self, ndim_obs, n_actions, hidden_sizes=(50, 50, 50)):
self.pi = policies.SoftmaxPolicy(model=links.MLP(
ndim_obs, n_actions, hidden_sizes, nonlinearity=F.tanh))
self.v = links.MLP(ndim_obs,
1,
hidden_sizes=hidden_sizes,
nonlinearity=F.tanh)
super().__init__(self.pi, self.v)
def __init__(self,
obs_size,
action_space,
n_hidden_layers=2,
n_hidden_channels=64,
bound_mean=None,
normalize_obs=None):
assert bound_mean in [False, True]
assert normalize_obs in [False, True]
super().__init__()
hidden_sizes = (n_hidden_channels, ) * n_hidden_layers
self.normalize_obs = normalize_obs
with self.init_scope():
self.pi = policies.FCGaussianPolicyWithStateIndependentCovariance(
obs_size,
action_space.low.size,
n_hidden_layers,
n_hidden_channels,
var_type='diagonal',
nonlinearity=F.tanh,
bound_mean=bound_mean,
min_action=action_space.low,
max_action=action_space.high,
mean_wscale=1e-2)
self.v = links.MLP(obs_size, 1, hidden_sizes=hidden_sizes)
if self.normalize_obs:
self.obs_filter = links.EmpiricalNormalization(shape=obs_size)
def __init__(self, n_in_elements, n_actions, explor_rate=0.0):
'''
Q値の範囲が報酬体系によって負の値をとる場合、F.reluは負の値をとれないので、学習に適さない。
活性化関数は、負の値も取ることが可能なものを選択する必要がある。
例えば、F.leaky_relu等。勾配消失問題を考えると、これが良い感じ。
n_size_twn_status: TWN自身の状態
num_ray_out: TWNセンサーをCNN層で処理した結果得られるデータ数のタプル (CNN層の出力要素数, CNN層の出力チャンネル数)
n_size_eb_status: TWNセンサーでとらえた、EBの情報
n_actions: 離散アクション空間
explor_rate=0.0: 探索行動比率（現時点で未使用）
'''
super().__init__()
with self.init_scope():
self.ml5 = links.MLP(
n_in_elements, n_actions,
(
n_in_elements*2,
int(n_in_elements*1.8),
int(n_in_elements*1.5),
int(n_in_elements*1.2),
n_in_elements,
int(n_in_elements*0.8),
(n_in_elements*2)//3,
(n_in_elements//2)*n_actions
),
nonlinearity=F.leaky_relu)
self.explor_rate = explor_rate
self.debug_info = None
def __init__(self, obs_space, action_space, out_size=1, gpu=-1):
hidden_sizes = (64, 64)
self.reward_net = links.MLP(obs_space + action_space,
out_size,
hidden_sizes=hidden_sizes)
self.value_net = links.MLP(obs_space,
out_size,
hidden_sizes=hidden_sizes)
if gpu >= 0:
self.reward_net.to_gpu(gpu)
self.value_net.to_gpu(gpu)
self.reward_optimizer = chainer.optimizers.Adam()
self.reward_optimizer.setup(self.reward_net)
self.value_optimizer = chainer.optimizers.Adam()
self.value_optimizer.setup(self.value_net)
def __init__(self,
n_size_twn_status,
num_ray,
n_size_eb_status,
n_actions,
explor_rate=0.0):
self.n_size_twn_status = n_size_twn_status
self.num_ray = num_ray
self.n_size_eb_status = n_size_eb_status
self.num_history = 1
self.n_clasfy_ray = 16
self.in_channel_1st = 1
out_channel_1st = 16
filter_size_1st = 5
slide_size_1st = 1
self.pooling_size_1st = 2
out_channel_2nd = 64
filter_size_2nd = 3
slide_size_2nd = 1
self.pooling_size_2nd = 4
self.num_of_out_elements_1st = self.calc_num_out_elements1D(
self.num_ray, self.in_channel_1st, out_channel_1st,
filter_size_1st, slide_size_1st, self.pooling_size_1st)
self.num_of_out_elements_2nd = self.calc_num_out_elements1D(
self.num_of_out_elements_1st, out_channel_1st, out_channel_2nd,
filter_size_2nd, slide_size_2nd, self.pooling_size_2nd)
#self.logger.info('1st out: {} 2nd out: {} n_actions'.format(self.num_of_out_elements_1st, self.num_of_out_elements_2nd, n_actions))
print('1st out: {} 2nd out: {} n_actions: {}'.format(
self.num_of_out_elements_1st, self.num_of_out_elements_2nd,
n_actions))
super().__init__()
with self.init_scope():
self.conv1 = L.ConvolutionND(
1, self.in_channel_1st, out_channel_1st,
filter_size_1st) # 1層目の畳み込み層（チャンネル数は16）
self.conv2 = L.ConvolutionND(
1, out_channel_1st, out_channel_2nd,
filter_size_2nd) # 2層目の畳み込み層（チャンネル数は64）
self.l3 = L.Linear(self.num_of_out_elements_2nd *
out_channel_2nd, self.n_clasfy_ray) #クラス分類用
aaa = n_size_twn_status + self.n_clasfy_ray + n_size_eb_status
self.l4 = L.Linear(aaa, aaa) #クラス分類用
self.ml5 = links.MLP(aaa,
n_actions,
(aaa * 3, aaa * 2,
(aaa // 2) * 3, aaa, aaa // 2),
nonlinearity=F.leaky_relu)
self.explor_rate = explor_rate
self.debug_info = None
def __init__(self, observation_dim, action_dim, hidden_sizes, loss_type='gan', gpu=-1):
self.model = links.MLP(observation_dim + action_dim, 1, hidden_sizes=hidden_sizes, nonlinearity=F.leaky_relu)
if gpu >= 0: self.model.to_gpu(gpu)
self.optimizer = chainer.optimizers.Adam(alpha=1e-5, eps=1e-5) # should alpha be somewhat higher?
self.optimizer.setup(self.model)
self.loss_type = loss_type
self.loss = None
def __init__(self,
n_in_elements,
n_original_input,
n_actions,
explor_rate=0.0):
'''
Q値の範囲が報酬体系によって負の値をとる場合、F.reluは負の値をとれないので、学習に適さない。
活性化関数は、負の値も取ることが可能なものを選択する必要がある。
例えば、F.leaky_relu等。勾配消失問題を考えると、これが良い感じ。
n_size_twn_status: TWN自身の状態
num_ray_out: TWNセンサーをCNN層で処理した結果得られるデータ数のタプル (CNN層の出力要素数, CNN層の出力チャンネル数)
n_size_eb_status: TWNセンサーでとらえた、EBの情報
n_actions: 離散アクション空間
explor_rate=0.0: 探索行動比率（現時点で未使用）
'''
super().__init__()
self.num_noise_roots = np.array([
n_original_input // 8, n_original_input // 8,
n_original_input // 8, n_original_input // 8
])
with self.init_scope():
self.l4 = links.MLP(
n_in_elements + np.sum(self.num_noise_roots),
n_original_input // 2, (n_original_input, ),
nonlinearity=F.leaky_relu)
self.action_links_list = []
for i in range(n_actions):
action_links = links.MLP(n_original_input // 2,
1, (n_original_input // 4, ),
nonlinearity=F.leaky_relu)
self.action_links_list.append(action_links)
self.explor_rate = explor_rate
self.debug_info = None
def __init__(self,
input_dim,
hidden_sizes=(64, 64, 64),
loss_type='wgangp',
gpu=-1):
self.model = links.MLP(input_dim, 1, hidden_sizes=hidden_sizes)
if gpu >= 0:
self.model.to_gpu(gpu)
self.optimizer = chainer.optimizers.Adam(alpha=1e-5, eps=1e-5)
self.optimizer.setup(self.model)
self.loss_type = loss_type
self.loss = None
def __init__(self, ndim_obs, n_actions, hidden_sizes=(64, 64)):
self.pi = policies.MellowmaxPolicy(
model=links.MLP(ndim_obs, n_actions, hidden_sizes))
self.v = links.MLP(ndim_obs, 1, hidden_sizes=hidden_sizes)
super().__init__(self.pi, self.v)