ReDUCE: Reformulation of Mixed Integer Programs Using Data from Unsupervised Clusters for Learning Efficient Strategies

Abstract

Mixed integer bilinear programs (MIBLPs) offer tools to resolve robotics motion planning problems with orthogonal rotation matrices or static moment balance, but require long solving times. Recent work utilizing data-driven methods has shown potential to overcome this issue allowing for applications on larger scale problems. To solve mixed-integer bilinear programs online with data-driven methods, several reformulations exist including mathematical programming with complementary constraints (MPCC), and mixed-integer programming (MIP). In this work, we compare the data-driven performances of various MIBLP reformulations using a book placement problem that has discrete configuration switches and bilinear constraints. The success rate, cost, and solving time are compared along with non-data-driven methods. Our results demonstrate the advantage of using data-driven methods to accelerate the solving speed of MIBLPs, and provide references for users to choose the suitable re-formulation.

Collaboration and Task Planning of Turtle-Inspired Multiple Amphibious Spherical Robots

“…Our future work will be showing that the proposed ADMM formulation can deal with long horizon tasks such as inserting multiple items in sequence. When the problem scales up, current ADMM formulation is expected to perform worse, hence data-driven methods such as [42] will be helpful.…”

“…For instance, in their work, [17] utilized graph neural networks to acquire heuristics, while [18] employed reinforcement learning to identify effective cutting planes. Alternatively, data collection can be employed to learn and address specific problems or tasks, as demonstrated in studies like [19]- [21]. This process effectively transforms into a classification problem, assigning a unique label to each strategy.…”