The Machine Learning Area at Microsoft Research Asia pushes the frontier of machine learning from the perspectives of theory, algorithms, and applications. Our research interests cover deep learning, ...

而且他们附加了一个 KL Penalty (惩罚项, 不懂的同学搜一下 KL divergence), 简单来说, 如果 new Policy 和 old Policy 差太多, 那 KL divergence 也越大, 我们不希望 new Policy 比 old Policy 差太多, 如果会差太多, 就相当于用了一个大的 Learning rate, 这样是不好的, 难收敛.