Learning for Constrained Optimization: Identifying Optimal Active Constraint Sets

TL;DR

This paper introduces a streaming learning algorithm that efficiently identifies relevant active constraint sets in optimization problems, reducing computational costs and enabling real-time decision making in systems like power flow management.

Contribution

The paper presents a novel streaming algorithm with theoretical guarantees for learning relevant active sets without assumptions on problem structure, applicable to large-scale constrained optimization.

Findings

Few active sets are relevant in practice, simplifying the learning task.

The algorithm converges quickly for problems with small relevant active sets.

Experiments on Optimal Power Flow demonstrate practical effectiveness.

Abstract

In many engineered systems, optimization is used for decision making at time-scales ranging from real-time operation to long-term planning. This process often involves solving similar optimization problems over and over again with slightly modified input parameters, often under tight latency requirements. We consider the problem of using the information available through this repeated solution process to directly learn a model of the optimal solution as a function of the input parameters, thus reducing the need to solve computationally expensive large-scale parametric programs in real time. Our proposed method is based on learning relevant sets of active constraints, from which the optimal solution can be obtained efficiently. Using active sets as features preserves information about the physics of the system, enables interpretable models, accounts for relevant safety constraints, and is easy to represent and encode. However, the total number of active sets is also very large, as it grows exponentially with system size. The key contribution of this paper is a streaming algorithm that learns the relevant active sets from training samples consisting of the input parameters and the corresponding optimal solution, without any assumptions on the problem structure. The algorithm comes with theoretical performance guarantees, and is known to converge fast for problem instances with a small number of relevant active sets. It can thus be used to establish the practicability of the learning method. Through extensive experiments on the Optimal Power Flow problem, we observe that often only a few active sets are relevant in practice, suggesting that the active sets is the appropriate level of abstraction for a learning algorithm to target.