About the EventThe emphasis of my research is on cognitive architecture, where the goal is
to develop the fixed processes, memories, representations, and interfaces
that support end-to-end behavior in intelligent systems. In this talk I will
present a case study of how multiple learning mechanisms (chunking and
reinforcement learning) are integrated in Soar, and how they combine to
produce novel learning in a multiplayer dice game where uncertainty is
rampant. Many AI game systems emphasize either complex evaluation functions
or learning by experience. In this talk, I will demonstrate how it is
straightforward to encode symbolic heuristics and opponent modeling as an
evaluation function in a Soar agent. I will also demonstrate how that
complex processing can be automatically compiled by chunking into selection
rules, which are tuned by reinforcement learning leading to significant
improvements in performance. This work may provide insight into the origins
of value functions in RL for large state spaces, and as to how those value
functions can be initialized and then tuned by experience. This research is
performed using existing mechanisms in Soar, without modification. |
