Search for: All records

We introduce CODD, a system for solving combinatorial optimization problems using decision diagram technology. Problems are represented as state-based dynamic programming models using the CODD language specification. The model specification is used to automatically compile relaxed and restricted decision diagrams that are embedded inside a branch-and-bound search process. We introduce abstractions that allow us to generically implement the solver components while maintaining overall execution efficiency. We demonstrate the functionality of CODD on a variety of combinatorial optimization problems and compare its performance to other state-based solvers as well as integer programming and constraint programming solvers. CODD provides competitive results and can outperform the other solvers, sometimes by orders of magnitude. 

For vehicle routing problems, strong dual bounds on the optimal value are needed to develop scalable exact algorithms as well as to evaluate the performance of heuristics. In this work, we propose an iterative algorithm to compute dual bounds motivated by connections between decision diagrams and dynamic programming models used for pricing in branch-and-cut-and-price algorithms. We apply techniques from the decision diagram literature to generate and strengthen novel route relaxations for obtaining dual bounds without using column generation. Our approach is generic and can be applied to various vehicle routing problems in which corresponding dynamic programming models are available. We apply our framework to the traveling salesman with drone problem and show that it produces dual bounds competitive to those from the state of the art. Applied to larger problem instances in which the state-of-the-art approach does not scale, our method outperforms other bounding techniques from the literature.

Funding: This work was supported by the National Science Foundation [Grant 1918102] and the Office of Naval Research [Grants N00014-18-1-2129 and N00014-21-1-2240].

Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2021.0170 .

 

Many real-world structured prediction problems need machine learning to capture data distribution and constraint reasoning to ensure structure validity. Nevertheless, constrained structured prediction is still limited in real-world applications because of the lack of tools to bridge constraint satisfaction and machine learning. In this paper, we propose COnstraint REasoning embedded Structured Prediction (Core-Sp), a scalable constraint reasoning and machine learning integrated approach for learning over structured domains. We propose to embed decision diagrams, a popular constraint reasoning tool, as a fully-differentiable module into deep neural networks for structured prediction. We also propose an iterative search algorithm to automate the searching process of the best Core-Sp structure. We evaluate Core-Sp on three applications: vehicle dispatching service planning, if-then program synthesis, and text2SQL generation. The proposed Core-Sp module demonstrates superior performance over state-of-the-art approaches in all three applications. The structures generated with Core-Sp satisfy 100% of the constraints when using exact decision diagrams. In addition, Core-Sp boosts learning performance by reducing the modeling space via constraint satisfaction. 

Haddock, introduced in [R. Gentzel et al., 2020], is a declarative language and architecture for the specification and the implementation of multi-valued decision diagrams. It relies on a labeled transition system to specify and compose individual constraints into a propagator with filtering capabilities that automatically deliver the expected level of filtering. Yet, the operational potency of the filtering algorithms strongly correlate with heuristics for carrying out refinements of the diagrams. This paper considers how to empower Haddock users with the ability to unobtrusively specify various such heuristics and derive the computational benefits of exerting fine-grained control over the refinement process. 

We present an exact algorithm for graph coloring and maximum clique problems based on SAT technology. It relies on four sub-algorithms that alternatingly compute cliques of larger size and colorings with fewer colors. We show how these techniques can mutually help each other: larger cliques facilitate finding smaller colorings, which in turn can boost finding larger cliques. We evaluate our approach on the DIMACS graph coloring suite. For finding maximum cliques, we show that our algorithm can improve the state-of-the-art MaxSAT-based solver IncMaxCLQ, and for the graph coloring problem, we close two open instances, decrease two upper bounds, and increase one lower bound. 

Relaxed decision diagrams have been successfully applied to solve combinatorial optimization problems, but their performance is known to strongly depend on the variable ordering. We propose a portfolio approach to selecting the best ordering among a set of alternatives. We consider several different portfolio mechanisms: a static uniform time-sharing portfolio, an offline predictive model of the single best algorithm using classifiers, a low-knowledge algorithm selection, and a dynamic online time allocator. As a case study, we compare and contrast their performance on the graph coloring problem. We find that on this problem domain, the dynamic online time allocator provides the best overall performance.