Discovering faster matrix multiplication algorithms with reinforcement learning