Intelligent agents are becoming increasingly important in our society in applications as diverse as house cleaning robots, computer-controlled opponents in video games, unmanned aerial combat vehicles, entertainment robots, and autonomous explorers in outer space. However, the broader adoption of intelligent agents is often hindered by their limited adaptability to new tasks; when conditions change slightly, agents may quickly become confused. Additionally, a substantial engineering effort is required to design an agent for each new task. This paper presents an adaptable, general purpose intelligent agent toolkit based on reinforcement learning (RL), an approach with strong mathematical foundations and intriguing biological implications. RL algorithms are powerful because of their generality: agents simply receive a scalar reward value representing success or failure, which greatly simplifies the agent design process. Furthermore, these algorithms can be combined with other techniques (e.g., planning from a learned internal model) to improve learning efficiency. The design and implementation of an open source RL toolkit is presented here as a step towards the goal of general purpose agents. Experimental results show learning performance on several tasks, including two physical control problems.

This content is only available via PDF.
You do not currently have access to this content.