Algorithmic Policy Decomposition for Deep Reinforcement Learning

By Victor Kolev

Deep reinforcement learning is a highly prospective and rapidly developing field of artificial intelligence. It involves training machine agents to interact with their environment in order to accomplish some final goal. A very intuitive example of this is a game, where a player (agent) needs to take certain actions to progress to the next level. Beyond games, however, deep reinforcement learning can be applied to numerous real-world settings, such as logistics optimization, resource management, and autonomous driving.

One of the key challenges in deep reinforcement learning is that solutions are often very unstable and cannot be scaled beyond lab testing, since the real world is a highly complex environment. To address this issue, I apply a well-known paradigm of computer science — divide and conquer. The problem is split into subproblems, for which individual AI modules are trained. In this manner, much greater scalability can be achieved, and the complexity of a task is alleviated.

Leave a Reply