Non-player characters (NPCs) in video games are a common form of frustration for players because they generally provide no explanations for their actions or provide simplistic explanations using fixed scripts. Motivated by this, we consider a new design for agents that can learn about their environments, accomplish a range of goals, and explain what they are doing to a supervisor. We propose a framework for studying this type of agent, and compare it to existing reinforcement learning and self-motivated agent frameworks. We propose a novel design for an initial agent that acts within this framework. Finally, we describe an evaluation centered around the supervisor's satisfaction and understanding of the agent's behavior.

Goal reasoning (GR) has a bright future as a foundation for the research and development of intelligent agents. GR is the study of agents that can deliberate on and self-select their goals/objectives, which is a desirable capability for some applications of deliberative autonomy. While studied in diverse AI sub-communities for multiple applications, our group has focused on how GR can play a key role for controlling autonomous systems. Thus, its importance is rapidly growing and it merits increased attention, particularly from the perspective of research on AI safety. In this article, I introduce GR, briefly relate it to other AI topics, summarize some of our group’s work on GR foundations and emerging applications, and describe some current and future research directions.

Agents with incomplete environment models are likely to be surprised, and this represents an opportunity to learn. We investigate approaches for situated agents to detect surprises, discriminate among different forms of surprise, and hypothesize new models for the unknown events that surprised them. We instantiate these approaches in a new goal reasoning agent (named FoolMeTwice), investigate its performance in simulation studies, and report that it produces plans with significantly reduced execution cost in comparison to not learning models for surprising events.

One drawback with using plan recognition in adversarial games is that often players must commit to a plan before it is possible to infer the opponent's intentions. In such cases, it is valuable to couple plan recognition with plan repair, particularly in multi-agent domains where complete replanning is not computationally feasible. This paper presents a method for learning plan repair policies in real-time using Upper Confidence Bounds applied to Trees (UCT). We demonstrate how these policies can be coupled with plan recognition in an American football game (Rush 2008) to create an autonomous offensive team capable of responding to unexpected changes in defensive strategy. Our real-time version of UCT learns play modifications that result in a significantly higher average yardage and fewer interceptions than either the baseline game or domain-specific heuristics. Although it is possible to use the actual game simulator to measure reward offline, to execute UCT in real-time demands a different approach; here we describe two modules for reusing data from offline UCT searches to learn accurate state and reward estimators.