maddpg/common/distributions.py [157:173]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def kl(self, other):
        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
        a1 = other.logits - U.max(other.logits, axis=1, keepdims=True)
        ea0 = tf.exp(a0)
        ea1 = tf.exp(a1)
        z0 = U.sum(ea0, axis=1, keepdims=True)
        z1 = U.sum(ea1, axis=1, keepdims=True)
        p0 = ea0 / z0
        return U.sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=1)
    def entropy(self):
        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
        ea0 = tf.exp(a0)
        z0 = U.sum(ea0, axis=1, keepdims=True)
        p0 = ea0 / z0
        return U.sum(p0 * (tf.log(z0) - a0), axis=1)
    def sample(self):
        u = tf.random_uniform(tf.shape(self.logits))
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



maddpg/common/distributions.py [188:204]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def kl(self, other):
        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
        a1 = other.logits - U.max(other.logits, axis=1, keepdims=True)
        ea0 = tf.exp(a0)
        ea1 = tf.exp(a1)
        z0 = U.sum(ea0, axis=1, keepdims=True)
        z1 = U.sum(ea1, axis=1, keepdims=True)
        p0 = ea0 / z0
        return U.sum(p0 * (a0 - tf.log(z0) - a1 + tf.log(z1)), axis=1)
    def entropy(self):
        a0 = self.logits - U.max(self.logits, axis=1, keepdims=True)
        ea0 = tf.exp(a0)
        z0 = U.sum(ea0, axis=1, keepdims=True)
        p0 = ea0 / z0
        return U.sum(p0 * (tf.log(z0) - a0), axis=1)
    def sample(self):
        u = tf.random_uniform(tf.shape(self.logits))
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -