Learning human motion models

Abstract

My research is focused on using human navigation data in games and simulation to learn motion models from trajectory data. These motion models can be used to: 1) track the opponent's movement during periods of network occlusion; 2) learn combat tactics by demonstration; 3) guide the planning process when the goal is to intercept the opponent. A training set of example motion trajectories is used to learn two types of parameterized models: 1) a second order dynamical steering model or 2) the reward vector for a Markov Decision Process. Candidate paths from the model serve as the motion model in a set of particle filters for predicting the opponent's location at different time horizons. Incorporating the proposed motion models into game bots allows them to customizes their tactics for specific human players and function as more capable teammates and adversaries. Copyright © 2012, Association for the Advancement of Artificial Intelligence. All rights reserved.