Computational Goals, Values and Decision-Making

  • PDF / 512,124 Bytes
  • 9 Pages / 439.37 x 666.142 pts Page_size
  • 90 Downloads / 203 Views

DOWNLOAD

REPORT


Computational Goals, Values and Decision‑Making Louise A. Dennis1

© The Author(s) 2020

Abstract Considering the popular framing of an artificial intelligence as a rational agent that always seeks to maximise its expected utility, referred to as its goal, one of the features attributed to such rational agents is that they will never select an action which will change their goal. Therefore, if such an agent is to be friendly towards humanity, one argument goes, we must understand how to specify this friendliness in terms of a utility function. Wolfhart Totschnig (Fully Autonomous AI, Science and Engineering Ethics, 2020), argues in contrast that a fully autonomous agent will have the ability to change its utility function and will do so guided by its values. This commentary examines computational accounts of goals, values and decision-making. It rejects the idea that a rational agent will never select an action that changes its goal but also argues that an artificial intelligence is unlikely to be purely rational in terms of always acting to maximise a utility function. It nevertheless also challenges the idea that an agent which does not change its goal cannot be considered fully autonomous. It does agree that values are an important component of decision-making and explores a number of reasons why.

Fully Autonomous AI In Fully Autonomous AI, Wolfhart Totschnig (2020) argues that the use of the word autonomy in much of the debate around Artificial Intelligence is philosophically weak: it presupposes that any such system possesses a fixed final goal that cannot be changed. He considers a fully autonomous agent to be one that can change its goals and then examines the mechanisms by which such goal changes might be effected. He argues that an intelligent system’s ability to change its goals will be based on a complex and nuanced understanding of what those goals are and that this, in turn, will be controlled by the system’s values. He uses this analysis to critique the argument that weak autonomy provides an assurance that if we can but define the final goal appropriately we need not fear the * Louise A. Dennis [email protected] 1



Center for Autonomous Systems Technology, Department of Computer Science, University of Liverpool, Liverpool, UK

13

Vol.:(0123456789)



L. A. Dennis

possible development of a super-intelligent AI that is antithetical to humanity (Bostrom 2014; Yudkowsky 2001). In this commentary I respond to the article from a computer science perspective and will focus on how we can understand the notion of value computationally. I will discuss the relationship between some fixed measurable quantity (that is often understood as a goal in AI systems) and a more abstract notion of goals, and the conceptual gap that arises when the task of creating a complex computational system reduces to that of defining a utility function. I will survey both alternative computational descriptions of goals, and alternative proposals for computational decision-making in AI systems and discuss how they rela