TitleGuiding the Ass with Goal Motivation Weights
Publication TypeConference Proceedings
Year of Conference2015
AuthorsMuñoz-Avila, H, Wilson, MA, Aha, DW
Conference Name2015 Conference on Advances in Cognitive Systems
Date PublishedMay 2015
Conference LocationAtlanta, Georgia

Goal reasoning (GR) is the study of free agents; they can autonomously and dynamically deliberate on and select what goals/objectives to pursue (Cox, 2007; Muñoz-Avila et al., 2010; Klenk et al., 2013; Vattam et al., 2013; Roberts et al., 2014). Endowing agents with this capability is particularly appropriate when the domain in which they operate is complex (e.g., partially observable, dynamic, multiagent), preventing the anticipation of all possible states and the precise pre-encoding of contingent plans for those states. Most GR agents monitor and assess the current state with respect to potential expectation violations or motivation triggers (Coddington et al., 2005). This deliberation may result in selecting an alternative to the goal(s) currently being pursued, requiring a planner to generate a corresponding set of actions for a controller to schedule and execute. In this paper, we study domain-independent goal selection, extending a method that combines motivations which was implemented in the GR agent M-ARTUE (Wilson et al., 2013). In particular, we relax the assumption that all motivations contribute equally to goal selection, and investigate the relationship between domain properties and motivator contribution in a paradigmatic domain. We view this as a step towards a deeper understanding of how motivations affect agent performance. Future work includes automatically learning motivator weights.

NRL Publication Release Number: