On AI, Markets and Machine Learning

What is fun about artificial intelligence (AI) is that it is fundamentally constructive! Rather than theorizing about the behavior of an existing, say social system, or understanding the way in which decisions are currently being made, one gets to ask a profound question: how should a system for making intelligent decisions be designed? In multi-agent systems we’re often interested, in particular, in what happens when multiple participants, each with autonomy, selfinterest and potentially misaligned incentives come together. These systems involve people (frequently people and firms), as well as the use of AI to automate some parts of decision making. How should such a system be designed? In adopting this normative viewpoint, we follow a successful branch of economic theory that includes mechanism design, social choice and matching. But in pursuit of success, we must grapple with problems that are more complex than has been typical in economic theory, and solve these problems at scale. We want to attack difficult problems by using the methods of AI together with the methods of economic theory. In this talk, I will highlight the following themes from my own research, themes that emerge as one lifts nice ideas from economic theory and brings them to bear on the kinds of problems that are at the heart of modern AI research: 1. Preferences need to be elicited, and cannot be assumed to be known or easily represented. 2. Mechanisms don’t need to be centralized. Rather, we can use (incentive aligned) distributed optimization! 3. Many interesting problems are temporal and involve uncertainty, leading to rich challenges that have been largely untouched by economic theory. 4. Markets become algorithms, and market-based optimization is a powerful and increasingly real paradigm. 5. Scale becomes an opportunity, as large systems beget data, this data enabling new approaches to robust incentive alignment. 6. Computation becomes a tool— machine learning has taken automated mechanism design up to the frontier of knowledge from 35 years of auction theory. Appears in: Proc. of the 16th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2017), S. Das, E. Durfee, K. Larson, M. Winikoff (eds.), May 8–12, 2017, São Paulo, Brazil. Copyright c 2017, International Foundation for Autonomous Agents and Multiagent Systems (www. ifaamas. org). All rights reserved. The future will bring a tighter and ever more compelling integration of AI with markets. The human and societal interface will remain, and become ever more important as there is more that can be automated. The anticipated“agentmediated economy”is almost upon us, and holds much promise as long as we make sure that AIs represent our preferences and system designs capture our values as society. CCS Concepts •Information systems → Electronic commerce; •Theory of computation → Algorithmic mechanism design; Computational pricing and auctions; •Computing methodologies → Multi-agent systems; Supervised learning;

Authors

• David C. Parkes Harvard University DeepMind

Keywords

No keywords are indexed for this paper.