Daniel Kudenko
http://www-users.cs.york.ac.uk/~kudenko/
Wednesday 9th January 2013
Time: 4pm
B10 Basement Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Knowledge-Based Reinforcement Learning
Reinforcement learning (RL) is a highly popular machine learning technique, mainly due to its natural fit to the agent paradigm and the resulting wide application potential. Nevertheless, RL still suffers from scalability problems, which have prevented its successful use in many complex real-world domains.
For many real-world tasks, human expert knowledge is available. For example, human experts have developed heuristics that help them in planning and scheduling resources. However, this domain knowledge is often rough and incomplete. While not perfect, the knowledge can be used to guide a reinforcement learning agent and restrict its policy search space. In addition, the RL agent's experience can be used to revise the knowledge.
In my talk I will introduce the research area of Knowledge-based Reinforcement Learning, focusing on potential-based reward shaping as a way to incorporate domain knowledge into the RL process to speed up the learning, as well as improve the quality of the result. The presentation will touch upon single- as well as multi-agent learning. I will finish by raising questions on whether reward shaping could be a suitable model in neuroscience for the influence of high-level cognitive reasoning on low-level behaviour.