PolyNet: Learning Diverse Solution Strategies for Neural Combinatorial Optimization
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: neural combinatorial optimization, learning to optimize, reinforcement learning, routing problems
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: A novel approach for neural combinatorial optimization that uses a single-decoder model to learn multiple complementary solution strategies.
Abstract: In recent years, learning-based approaches have made remarkable strides in tackling combinatorial optimization problems. Reinforcement learning-based construction methods, in particular, have shown promise in producing high-quality solutions, often surpassing established operations research heuristics for simple routing problems. Nonetheless, inherent limitations, such as a lack of solution diversity and limited applicability to complex problems, have hindered their widespread adoption. This paper introduces PolyNet, a novel approach that uses a single-decoder model to learn complementary solution strategies for combinatorial optimization problems, allowing the rapid creation of diverse solutions for a given instance. Moreover, PolyNet's diversity mechanism enhances training exploration without relying on solution space symmetries, enabling it to effectively tackle more complex problems. We evaluate PolyNet on three combinatorial optimization problems of varying difficulty. Our comprehensive experiments consistently demonstrate significant improvements over state-of-the-art machine learning methods, both in terms of swift solution generation and extensive search.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5375
Loading