Learning and Decision-Making for Intention Reconciliation

Abstract

Rational, autonomous agents must be able to revise their commitments in the light of new opportunities. They must decide when to default on commitments to the group in order to commit to potentially more valuable outside offers. The SPIRE experimental system allows the study of intention reconciliation in team contexts. This paper presents a new framework for SPIRE that allows for mathematical specification and provides a basis for the study of learning. Analysis shows that a reactive policy can be expected to perform as well as more complex policies that look ahead. We present an algorithm for learning when to default on group commitments based solely on observed values of group-related tasks and discuss the applicability of this algorithm in settings where multiple agents may be learning.