Hybridized Assistance Games and Value Alignment

Abstract

A fairly recent proposal in the study of value alignment is the assistance game, in which initially unsure agents learn to maximize human preferences by observing human behaviour. Here, we propose that assistance game-based agents might benefit from being ”hybridized” with other AI techniques. To describe these hybridized systems, we first consider the advantages and disadvantages of assistance games, before considering in what ways a hybridized agent may work and how an assistance game-based agent with sufficient computational resources might be motivated to create a hybridized system by using other AI technique(s). To illustrate the beneficial effects of a hybridized system, we consider ways the effects of these systems might fulfill the requirements of trustworthy AI described by the European Union’s High Level Expert Group on Artificial Intelligence.