Intelligent Affect: Rational Decision Making for Socially Aligned Agents

Abstract

Affect Control Theory (ACT) is a mathematical model that makes accurate predictions about human behaviour across a wide range of settings. The predictions, which are derived from statistics about human actions and identities in real and laboratory environments, are shared prescriptive and affective behaviours that are believed to lead to solutions to everyday cooperative problems. A generalisation of ACT, called BayesACT, allows the principles of ACT to be used for human-interactive agents by combining a probabilistic version of the ACT dynamical model of affect with a utility function encoding external goals. Planning in BayesACT, which we address in this paper, then allows one to go beyond the affective prescription, and leads to the emergence of more complex interactions between ``cognitive'' reasoning and ``affective'' reasoning, such as deception leading to manipulation and altercasting. We use a continuous variant of a successful Monte-Carlo tree search planner (POMCP), which performs dynamic discretisation of the action and observation spaces while planning. We present results on two classic two-person social dilemmas, and show how reasoning about affect can produce some remarkably powerful, yet human-like, strategies in these games.