October 1, 2024

A Temporal Convolutional Neural Network (TCNN) Approach to Predicting Capacitated Lot-Sizing Solutions

Key Points

The TCNN approach predicts optimal decisions for capacitated lot sizing, and reduces solution time significantly.
This method demonstrates a reduction in CPLEX solution time by a factor of 14 while maintaining a 0.5% gap from optimality.
Using temporal convolutional neural networks, the framework excels over LSTM models, particularly for mixed-integer programs.
These findings highlight the potential of TCNN for enhancing efficiency in sequential decision-making problems.

Abstract

This study presents an innovative deep learning optimization framework designed for dynamic mixed-integer programs, featuring the integration of a temporal convolutional neural network (TCNN) with optimization. The framework is applied to predict optimal decisions for the Single-Item Capacitated Lot Sizing Problem (CLSP), where binary variables govern production setup within discrete periods, transforming it into a sequence labeling task. Our hybrid TCNN-optimization framework reduces the CPLEX solution time by a significant factor of 14, while deviating from optimality with a 0.5% gap. We also present a generalized LSTM modeling approach and compare the results with those obtained by TCNN. Our results show the dominance of the TCNN model over the LSTM model in terms of time improvement and performance, especially when using a generalized model that could handle different types of CLSP data, while LSTM may perform better than TCNN when trained on a separate data set. These findings highlight the potential of our hybrid TCNN-optimization framework as a promising solution to efficiently solve sequential decision-making problems in repetitive scenarios.

Demander à l'IA

Bookmark