Methods for Exploiting Past Experience Data in Reinforcement Learning

Recent advances in autonomy technology have promoted the widespread emergence of autonomous systems in various domains such as domestic robots, self-driving vehicles, and financial management agents. The technology developed in this invention, Fitted Q-Iteration with Complex Returns (CFQI), enables an autonomous system to learn how to perform a complex task from past experiences through reinforcement learning – learning from feedback without human intervention. Successful application of this technology will not only make it easier and faster to build general-purpose autonomous systems, but also enable an autonomous system to continuously improve its performance and adapt to new and dynamic environments.

Advantages:

Compared with competitive products without using our technology, products based on our technology possess the following advantages: A) Less dependent on human intervention (e.g., anticipating the operational environment and hard coding rules of operations and learning) in teaching a system how to perform a task, and therefore, less human cost, and faster to develop a system; B) Ability to adapt to new, uncertain, dynamic environment without human intervention (e.g., specifically instructed by a human expert to change its course of actions under a new condition); and C) Forever learning capability - the longer a system is deployed, the more experience it gets, and the better it performs. Compared with alterative RL processes, in particular, existing FQI methods, our technology possesses the following advantages: i) Better sample efficiency than FQI – achieving the same level of policy performance using significantly less samples; ii) Better computational efficiency than FQI – achieving the same level of policy performance with significantly less computation time; and iii) Better effectiveness than TFQI (an earlier extension to FQI) – much more reliable in producing improved policy performance over FQI given the same set of samples.

 

Binghamton University RB488

Patent Information: