A Unified Framework for Planning and Acting

Considerable effort has been put into hybrid behavior architectures, with separate components for long-term planning and short-term acting. These systems dominate the field today, to such an extent that it is often assumed that the fundamental tradeoff in real-time behavior is between reactive execution and deliberative inference. But this division is not justified. Deliberation is not always long-term, nor is reacting the same as acting. We believe we can build an agent that has one mechanism for both reactive and deliberative behavior, and for both planning and acting. This is possible because of three innovations: (1) A representation of planning as mental action, which obviates the need for separate representations and mechanisms for inference and acting. (2) The use of mathematics related to nonlinear dynamics to learn inferences and behaviors as attractors of the system, so that long-chain inferences and behaviors are not random, and so agents switch behaviors when necessary, but not continually. (3) The use of mathematics from percolation theory to ensure that the chains of inference and action produced are dominated neither by short, reactive chains, nor by long, deliberative chains, but by a power-law distribution that produces chains of all lengths.