67f74e592427d15578eae688f677952d8bd98d3a,tensorforce/core/distributions/categorical.py,Categorical,tf_parametrize,#Categorical#Any#Any#,80

Before Change

    def tf_parametrize(self, x, mask):
        epsilon = tf.constant(value=util.epsilon, dtype=util.tf_dtype(dtype="float"))
        shape = (-1,) + self.action_spec["shape"] + (self.action_spec["num_values"],)
        value_shape = (-1,) + self.action_spec["shape"] + (1,)

        // Deviations
        action_values = self.deviations.apply(x=x)
        action_values = tf.reshape(tensor=action_values, shape=shape)

After Change

        else:
            // Explicit states value and advantage-based action values
            states_value = self.state_value.apply(x=x)
            states_value = tf.reshape(tensor=states_value, shape=shape[:-1])
            action_values = tf.expand_dims(input=states_value, axis=-1) + action_values
            action_values -= tf.math.reduce_mean(input_tensor=action_values, axis=-1, keepdims=True)

        // TODO: before or after states_value?

Instances