Prof. Yi Mei

Prof. Yi Mei

Professor, Fellow of Engineering New Zealand, IEEE Senior Member

Transfer, Multitask, and Lifelong Machine Learning

Problems rarely come alone. A factory faces many related scheduling scenarios; a logistics company solves new variants of routing every day. Instead of learning each from scratch, my research reuses knowledge — through transfer learning, multitask learning and lifelong learning — so that solvers learn faster and generalise further.

30+papers in this theme
2award-winning papers
3papers with 100+ citations (OpenAlex)
ICLR2026 — lifelong learning for routing

Introduction

Three ways to reuse knowledge

Transfer learning reuses what was learned on a source task to speed up a target task. Multitask learning solves several related tasks together so they help each other. Lifelong learning faces a stream of tasks and must keep improving without forgetting earlier ones.

Transfer, multitask and lifelong learningTRANSFERMULTITASKLIFELONGsourcetargetlearn once, reuseon a new tasktask Atask Blearn related taskstogether, sharing knowledgeT1T2T3growing knowledge basekeep learning without forgetting
Figure 1. Transfer, multitask and lifelong learning.

My Contributions

With my students and collaborators, I have made the following contributions. The papers shown are selected automatically: award-winning papers, papers with 100+ citations, and papers in top venues. Click a title to read the paper, or list entry to see it in my publication list.

1 Knowledge transfer in genetic programming

We transfer useful sub-trees, individuals and search distributions between GP runs on related routing problems, and studied how reusable evolved rules are.

2 Evolutionary multitask learning

Multitask GP learns heuristics for several scenarios simultaneously, deciding which tasks are related and what to share.

Multitask GP with knowledge transferTASK Ae.g. minimise mean flowtimeTASK Be.g. minimise mean tardinesstransfer promising individualsshare building blockssurrogates rank rules cheaply
Figure 2. Multitask GP shares promising individuals and building blocks between tasks, and uses surrogates to save evaluations.

3 Lifelong learning for optimisation

When new problem variants keep arriving, a learner must adapt without forgetting. We consolidate past behaviour to keep performance on earlier tasks while learning new ones.

Lifelong learning for optimisationA STREAM OF OPTIMISATION TASKS ARRIVES OVER TIMECVRPVRP + time windowsVRP + pickupnew variant …Learner(neural solver / optimiser generator)memory of pastbehaviourconsolidate, don’t forgetgood on newand old tasks
Figure 3. Lifelong learning for optimisation: a learner faces a stream of routing variants and consolidates its past behaviour to avoid forgetting.

4 Generalising across problems

We improve how learned solvers and heuristics generalise to new sizes, variants and objectives, with parameter sharing, attention reshaping and curriculum learning.

Research Output

Papers per yearStacked bars of journal and conference papers per year.02462016: 0 journal2016: 1 conference120162017: 0 journal2017: 1 conference120172018: 0 journal2018: 1 conference120182019: 0 journal2019: 4 conference420192020: 0 journal2020: 5 conference520202021: 2 journal2021: 2 conference420212022: 2 journal2022: 1 conference320222023: 4 journal2023: 1 conference520232024: 2 journal2024: 1 conference320242025: 2 journal2025: 2 conference420252026: 0 journal2026: 3 conference32026Journal articles & booksConference papers
Figure 4. My papers in this theme per year (34 in total), identified by title keywords in my publication lists; papers can belong to more than one theme. The 2026 count covers papers published so far this year.

Research Themes

Full publication list →   ← Back to home