Toward a Model of Mind as a Laissez-Faire Economy of Idiots

A learning machine called \The Hayek Machine" is proposed and tested on a simulated Blocks World planning problem. Hayek learns a set of agents from reinforcement. The agents interact in a market economy. A price mechanism is proposed and seen to have three desirable eeects. First, the market price learns to estimate Hayek's future reward from using a given agent. Second, the market automatically selects the agent with highest estimate to act next. Third, new agents can enter the market if and only if they have greater expected utility than direct competitors. Hayek learns by gradual accretion of useful agents and elimination of poor ones,and by reenement of its price estimates. Many agents act in consort to solve problems. Our Blocks World (BW) problems, which involve discovering the abstract goal of copying a stack, and necessitate solving Towers of Hanoi-like problems, are far more complex than any BW problems previously addressed by a learning algorithm. Starting from tabula rasa, and with reinforcement only upon completion of an instance, Hayek learns to solve eight block problems. These are larger than can be solved by a simple handcrafted program, and of comparable complexity to those solvable by general purpose planners. Given intermediate reinforcement or simple features such as \top of stack", Hayek does much better. Given both, Hayek produces a set of agents which solves arbitrary Blocks World Instances, and does so eeciently on almost all such instances. My goals in studying Hayek are threefold. First, Hayek represents progress toward a useful program for learning to reason. Second, it is intended as a model of mind. It models how human-like mental capabilities can be autonomously broken into simple components nowhere invoking a homunculus. Third, it is a model of economics. Utilizing very limited agents and starting far from equilibrium, Hayek achieves a productive and stable economy. It thus gives insight as to why and how far mathematical economic predictions based on assumptions of computationally powerful and rational agents, and on convergence to equilibrium, can be expected to hold in the real world.