The Path Integral Bottleneck: Exploring the Control-Compute Tradeoff

Abstract: Executing a control sequence requires some computation effort. Intuitively, a high-effort, fine-grained computation should result in better control (e.g. lower cost), whereas little to no computation effort would lead to worse control. To quantify and explore the tradeoff between control performance and compute effort, we present the Path Integral Bottleneck (PIB), a fusion of the Path Integral (PI) optimal control and Information Bottleneck (IB) frameworks. Both frameworks provide flexible and probabilistic descriptions of control. The PI does not limit itself to a particular control law, and the IB is not bound to any specific state encoding. Combining the generality of both frameworks enables us to produce an analytical description of the control-compute tradeoff. We provide PIB formulations for both continuous and discrete random variables. With these formulations, we can plot a tradeoff curve between performance and computation effort for any given plant description and control cost function. Simulations of a cart-pole for both the continuous and discrete variable cases reveal fundamental control-compute tradeoffs, exposing regions where the task performance-per-compute is higher than others.