Decision markets implementations for human forecasters and multi-agent learning systems : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Computer Science at Massey University, Albany, New Zealand

Abstract

Mechanisms of collective decision-making are an increasingly important topic, given that relevant data and information are often distributed. Collective decision-making processes involve eliciting information from multiple agents, aggregating the information, and mapping the aggregated information to a decision. An obstacle to these processes is that information is often proprietary, held by self-interested agents, and sometimes even too sensitive to share. Decision markets are mechanisms for eliciting and aggregating such information into predictions for decision-making. A design for decision markets put forward by Chen, Kash, Ruberry, et al. uses prediction markets to elicit and aggregate predictions that are conditional to the available actions, and then uses a stochastic decision rule to determine, based on the aggregated forecasts, which action to select. The design is incentive-compatible and uses a decision scoring rule to evaluate and incentivise the self-interested agents for their forecasts.

The first part of this thesis (Chapter 2) describes a framework for security-based decision markets that allows agents to make predictions by trading assets. Security-based decision markets are designed to be user-friendly for participants familiar with trading in stock markets. For prediction markets, such a framework is well studied. For decision markets, my results show there are important differences between scoring rule based and securities-based implementation.

The second and third parts of this thesis (Chapters 3 and 4) investigate decision markets as mechanisms of collective decision-making for multi-agent learning problems, thus building a bridge between economic mechanisms and artificial intelligence. Chapter 3 provides a decision market based algorithm that allows a principal to train multiple autonomous agents with independent and identically distributed (iid) information to solve a contextual bandit problem. Simulation results demonstrate that the proposed multi-agent systems can achieve performance equivalent to a centralised counterpart without requiring direct access to the agents' iid information, which is necessary for the centralised counterpart.

Chapter 4 describes a set of mechanisms that allow avoiding stochastic decision rules to select actions based on aggregated forecasts. This is important because committing to a stochastic (i.e., randomising) decision rule means that sometimes suboptimal decisions have to be taken. The mechanisms outlined in this chapter require agents to collectively predict a proxy instead of conditional outcomes. Simulations show that the performance is as good as a Bayesian model with access to all distributed information.