Optimal control with moderation incentives

TL;DR

This paper introduces moderation incentives into optimal control problems, enabling smoother control solutions within the admissible region, and demonstrates their advantages over traditional quadratic costs through analytical and numerical examples.

Contribution

It develops a new class of moderation incentives, including elliptical forms, that improve control smoothness and solution existence in optimal control problems.

Findings

Elliptical incentives produce smooth interior controls.

Traditional quadratic costs may lack solutions for certain durations.

Moderation incentives can simplify numerical solution methods.

Abstract

A purely state-dependent cost function can be modified by introducing a control-dependent term rewarding submaximal control utilization. A moderation incentive is identically zero on the boundary of the admissible control region and non-negative on the interior; it is bounded above by the infimum of the state-dependent cost function, so that the instantaneous total cost is always non-negative. The conservation law determined by the Maximum Principle, in combination with the condition that the moderation incentive equal zero on the boundary of the admissible control region, plays a crucial role in the analysis; in some situations, the initial and final values of the auxiliary variable are uniquely determined by the condition that the conserved quantity equal zero along a solution of the arbitrary duration synthesis problem. Use of an alternate system of evolution equations, parametrized by the auxiliary variable, for one-degree of freedom controlled acceleration systems, can significantly simplify numerical searches for solutions of the arbitrary duration synthesis problem. A one-parameter family of 'elliptical' moderation incentives is introduced; the behavior of the well-known quadratic control cost and its incentive analog is compared to that of the elliptical incentives in two simple controlled acceleration examples. The elliptical incentives yield smooth solutions with controls remaining in the interior of the admissible region, while the quadratic incentive allows piecewise smooth solutions with controls moving on and off the boundary of the admissible region; in these examples, the arbitrary duration synthesis problem for the traditional quadratic control cost has no solution; the total cost is a monotonically decreasing function of the duration.