Economic reasoning and artificial intelligence

Abstract

The field of artificial intelligence (AI) strives to build rational agents capable of perceiving the world around them and taking actions to advance specified goals. Put another way, AI researchers aim to construct a synthetic homo economicus, the mythical perfectly rational agent of neoclassical economics. We review progress toward creating this new species of machine, machina economicus, and discuss some challenges in designing AIs that can reason effectively in economic contexts. Supposing that AI succeeds in this quest, or at least comes close enough that it is useful to think about AIs in rationalistic terms, we ask how to design the rules of interaction in multi-agent systems that come to represent an economy of AIs.Theories of normative design from economics may prove more relevant for artificial agents than human agents, with AIs that better respect idealized assumptions of rationality than people, interacting through novel rules and incentive systems quite distinct from those tailored for people. Economics models the behavior of people, firms, and other decision-makers as a means to understand how these decisions shape the pattern of activities that produce value and ultimately satisfy (or fail to satisfy) human needs and desires. In this enterprise, the field classically starts from an assumption that actors behave rationally-that is, their decisions are the best possible given their available actions, their preferences, and their beliefs about the outcomes of these actions. Economics is drawn to rational decision models because they directly connect choices and values in a mathematically precise manner. Critics argue that the field studies a mythical species, homo economicus ("economic man") and produces theories with limited applicability to how real humans behave. Defenders acknowledge that rationality is an idealization but counter that the abstraction supports powerful analysis, which is often quite predictive of people's behavior (as individuals or in aggregate). Even if not perfectly accurate representations, rational models also allow preferences to be estimated from observed actions and build understanding that can usefully inform policy.Artificial intelligence (AI) research is likewise drawn to rationality concepts, because they provide an ideal for the computational artifacts it seeks to create. Core to the modern conception of AI is the idea of designing agents: entities that perceive the world and act in it (1). The quality of an AI design is judged by how well the agent's actions advance specified goals, conditioned on the perceptions observed. This coherence among perceptions, actions, and goals is the essence of rationality. If we represent goals in terms of preference over outcomes, and conceive perception and action within the framework of decisionmaking under uncertainty, then the AI agent's situation aligns squarely with the standard economic paradigm of rational choice. Thus, the AI designer's task is to build rational agents, or agents that best approximate rationality given the l...

Bounded Rationality, Abstraction, and Hierarchical Decision-Making: An Information-Theoretic Optimality Principle

“…The fundamental trade-off between large expected utility and low computational cost appears in many domains such as machine learning, AI, economics, computational biology or neuroscience, and many solutions, such as heuristics, samplingbased approaches, and model-based approximation schemes, exist (Gershman et al, 2015;Jordan and Mitchell, 2015;Parkes and Wellman, 2015). One of the exciting prospects of such an approach is that it might provide a common ground for research-questions from artificial intelligence and neuroscience, thus partially unifying the two fields that share common origins but have drifted apart over the last decades (Gershman et al, 2015).…”

“…One way to deal with such problems is to study optimal decisionmaking with information-processing constraints. Following the pioneering work of Simon (1955Simon ( , 1972 on bounded rationality, decision-making with limited information-processing resources has been studied extensively in psychology (Gigerenzer and Todd, 1999;Camerer, 2003;Brighton, 2009), economics (McKelvey andPalfrey, 1995;Rubinstein, 1998;Kahneman, 2003;Parkes and Wellman, 2015), political science (Jones, 2003), industrial organization (Spiegler, 2011), cognitive science (Howes et al, 2009;Janssen et al, 2011), computer science, and artificial intelligence research (Horvitz, 1988;Lipman, 1995;Russell, 1995;Russell and Subramanian, 1995;Russell and Norvig, 2002;Lewis et al, 2014). Conceptually, the approaches differ widely ranging from heuristics (Tversky and Kahneman, 1974;Gigerenzer and Todd, 1999;Gigerenzer and Brighton, 2009;Burns et al, 2013) to approximate statistical inference schemes (Levy et al, 2009;Vul et al, 2009Vul et al, , 2014Sanborn et al, 2010;Tenenbaum et al, 2011;Fox and Roberts, 2012;Lieder et al, 2012).…”

Towards Scalable Governance: Sensemaking and Cooperation in the Age of Social Media

“…Here, again, we face issues of strategic manipulation (Faliszewski and Procaccia 2010), impossibility of various kinds of fair resource allocations (Procaccia 2013), and so on. These are all problems that researchers in the algorithmic game theory community are exploring in depth (Nisan et al 2007;Chen et al 2013;Parkes and Wellman 2015). …”

Grammars for Games: A Gradient-Based, Game-Theoretic Framework for Optimization in Deep Learning

“…Although game theory was developed to model human interactions (von Neumann and Morgenstern, 1944), it has been pointed out that it may be more directly applicable to interacting populations of algorithms, the so-called machina economicus (Lay and Barbu, 2010;Abernethy and Frongillo, 2011;Storkey, 2011;Frongillo and Reid, 2015;Parkes and Wellman, 2015;Syrgkanis et al, 2015). This paper goes one step further to propose that games played over first-order communication protocols are a key component of the foundations of deep learning.…”