Title: Machine Learning of Human Behavior for Opponents in Video Games

Speaker: Nathan Fabian, Sandia National Laboratories

Date/Time: Monday, February 18, 2008 from 3-4 pm

Location: CSRI Building, Room 90 (Sandia NM)

Brief Abstract: I present a technique for creating video game agents using machine learning to clone the behavior of human trainers playing the game.  I compare the results of C4.5, a traditional attribute-value learning algorithm against first-order, relational learning algorithms:  FOIL, PROGOL, and PROXIMITY.  I establish a synthetic framework in which to analyze various attributes of behavior learning, and I will argue that although FOIL is more fragile as a result of being a greedy search, it is the most pragmatic option of the three.  Finally I show a comparison of results of C4.5 and FOIL on human trainers.

CSRI POC: David Rogers, 844-5323