the network for evaluate values of state-action pairs: Q(s,a)
| 116 | |
| 117 | |
| 118 | class SoftQNetwork(Model): |
| 119 | """ the network for evaluate values of state-action pairs: Q(s,a) """ |
| 120 | |
| 121 | def __init__(self, num_inputs, num_actions, hidden_dim, init_w=3e-3): |
| 122 | super(SoftQNetwork, self).__init__() |
| 123 | input_dim = num_inputs + num_actions |
| 124 | w_init = tf.keras.initializers.glorot_normal( |
| 125 | seed=None |
| 126 | ) # glorot initialization is better than uniform in practice |
| 127 | # w_init = tf.random_uniform_initializer(-init_w, init_w) |
| 128 | |
| 129 | self.linear1 = Dense(n_units=hidden_dim, act=tf.nn.relu, W_init=w_init, in_channels=input_dim, name='q1') |
| 130 | self.linear2 = Dense(n_units=hidden_dim, act=tf.nn.relu, W_init=w_init, in_channels=hidden_dim, name='q2') |
| 131 | self.linear3 = Dense(n_units=1, W_init=w_init, in_channels=hidden_dim, name='q3') |
| 132 | |
| 133 | def forward(self, input): |
| 134 | x = self.linear1(input) |
| 135 | x = self.linear2(x) |
| 136 | x = self.linear3(x) |
| 137 | return x |
| 138 | |
| 139 | |
| 140 | class PolicyNetwork(Model): |
no outgoing calls
no test coverage detected
searching dependent graphs…