Tightly Bounded Polynomials via Flexible Discretizations for Dynamic Optimization Problems (2403.07707v2)

Abstract: Polynomials are widely used to represent the trajectories of states and/or inputs. It has been shown that a polynomial can be bounded by its coefficients, when expressed in the Bernstein basis. However, in general, the bounds provided by the Bernstein coefficients are not tight. We propose a method for obtaining numerical solutions to dynamic optimization problems, where a flexible discretization is used to achieve tight polynomial bounds. The proposed method is used to solve a constrained cart-pole swing-up optimal control problem. The flexible discretization eliminates the conservatism of the Bernstein bounds and enables a lower cost, in comparison with non-flexible discretizations. A theoretical result on obtaining tight polynomial bounds with a finite discretization is presented. In some applications with linear dynamics, the non-convexity introduced by the flexible discretization may be a drawback.