Partial planning for situated agents based on active logic
Slawek Nowaczyk, AI@CS, LU
AI seminar, November 30th, 2006

Abstract:

In this seminar I will be talking about the research I am doing for my
PhD, which concerns machine learning for rational agents, more
specifically one utilising rich knowledge representations. In
particular, my main interests lie in designing and applying learning
algorithms to problems in which available knowledge is highly
structured, and in exploring the inherent structure to achieve more
human-like behaviour.

On of the domains I am especially interested in are rational agents
situated in a dynamic or unknown world. In this seminar I will talk
about ideas for combining partial planning with learning, which I hope
will allow agents to make rational decisions in face of partial
information without the need to analyse every possible situation.

The basic idea is to allow an agent to create short, conditional plans
of actions for the nearest future. By using learning techniques to
choose (one of) the best of such plans, executing them and observing
their results, agents can achieve acceptable performance without the
need to reason about all possible outcomes and hypothetical
situations.