LIMIP: Lifelong Learning to Solve Mixed Integer Programs

Mixed Integer programs (MIPs) are typically solved by the Branch-and-Bound algorithm. Recently, Learning to imitate fast approximations of expert strong branching heuristic has gained attention due its success in reducing running time for solving MIPs. However, existing learning-to-branch methods assume that entire training data is available a single session training. This assumption often not true, and if supplied continual fashion over time, techniques suffer from catastrophic forgetting. In this work, we study hitherto unexplored paradigm Lifelong Branch on Programs. To mitigate forgetting, propose LIMIP, which powered idea modeling an MIP instance form bipartite graph, map embedding space using Graph Attention Network. rich avoids forgetting through application knowledge distillation elastic weight consolidation, wherein learn parameters key towards retaining efficacy therefore protected significant drift. We evaluate LIMIP series NP-hard problems establish comparison baselines, up 50% better when confronted with lifelong learning.