Prof. Yi Mei

Prof. Yi Mei

Professor, Fellow of Engineering New Zealand, IEEE Senior Member

Genetic Programming for Real-time Optimisation and Decision-Making

Real-world systems change constantly: jobs arrive, machines break down, demands and travel times are uncertain. Re-optimising after every change is too slow. My research uses genetic programming (GP) to automatically learn decision rules that make high-quality decisions in real time — in microseconds, with formulas people can read.

160+papers in this theme
6award-winning papers
10papers with 100+ citations (OpenAlex)
1stbook on hyper-heuristics for production scheduling

Introduction

Decisions that cannot wait

In a dynamic job shop, every time a machine becomes idle it must choose which waiting job to process next; in uncertain arc routing, a vehicle must choose its next street when it finishes a task. These decisions must be made immediately, using only the current state.

Real-time decision pointsDynamic events arrive over time; at each decision point the evolved rule decides instantly.job arrivesmachine idle → rule ranks queuemachine breaks downurgent job arrivesvehicle idle → rule picks next tasktime
Figure 1. Dynamic events arrive over time. At each decision point (filled circles) an evolved rule ranks the options instantly.
A dispatching rule as a GP tree: PT + WINQ / NOR+PT÷WINQNORfunctionterminal (job / machine feature)
Figure 2. A scheduling rule as a GP tree, computing priority PT + WINQ / NOR from processing time, work in the next queue and number of operations remaining.

Why genetic programming?

  • Learns the rule itself — GP searches over programs, discovering new rule structures rather than tuning parameters.
  • Fast at decision time — evaluating a formula takes microseconds.
  • Readable — experts can inspect and trust the evolved rule.
  • Trained in simulation — learned offline on many scenarios, reused online.
GP hyper-heuristic workflowOffline, candidate rules are evolved in simulation; online, the best rule makes real-time decisions.OFFLINE · learn the heuristicONLINE · use the heuristic1. Population ofcandidate rules (GP trees)2. Evaluate each rule insimulated dynamic scenarios3. Select & vary(crossover, mutation)Best evolved rule(compact, interpretable)Real-time decisions: rankwaiting jobs / vehicles / tasks
Figure 3. A GP hyper-heuristic learns the rule offline in simulation; after a number of generations the best rule is deployed online.

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 Scheduling heuristics for dynamic job shops

Our core line of work makes GP hyper-heuristics effective for dynamic (flexible) job shop scheduling: evolving routing and sequencing rules together, selecting the most useful features, and combining rules into ensembles. It is summarised in our Springer book and surveys.

2 Routing policies under uncertainty

For arc and vehicle routing with uncertain demands and costs, we evolve routing policies that let each vehicle decide its next task on the fly, including vehicle collaboration, niching and multi-objective trade-offs.

3 New program representations

We introduced linear and grammar-guided linear GP to scheduling, enabling rules with reusable sub-expressions and flow-control logic.

Tree-based and linear representationsTREE-BASED GP+PT÷WINQNOR≡LINEAR GPr1 ← WINQ ÷ NORr0 ← PT + r1return r0registers let sub-results be reused (a graph, not a tree)
Figure 4. The same rule in tree-based GP and in linear GP, where registers store intermediate results for reuse.

4 Learning faster with surrogates

Evaluating a rule needs many simulations. Our surrogate models and multi-fidelity evaluation estimate rule quality cheaply, cutting training time substantially.

5 Real-time decisions beyond the factory

The same approach powers real-time decisions in many domains.

Traffic signal controlCloud & fog workflow schedulingAir traffic flow managementYard crane schedulingEmergency medical dispatchMulti-agent path planning

Multitask learning, explainability and links with reinforcement learning in GP are covered under the other themes below.

Research Output

Papers per yearStacked bars of journal and conference papers per year.051015202014: 0 journal2014: 2 conference220142015: 0 journal2015: 1 conference120152016: 0 journal2016: 6 conference620162017: 2 journal2017: 11 conference1320172018: 1 journal2018: 12 conference1320182019: 1 journal2019: 14 conference1520192020: 3 journal2020: 10 conference1320202021: 6 journal2021: 11 conference1720212022: 6 journal2022: 14 conference2020222023: 7 journal2023: 7 conference1420232024: 11 journal2024: 8 conference1920242025: 10 journal2025: 10 conference2020252026: 9 journal2026: 5 conference142026Journal articles & booksConference papers
Figure 5. My papers in this theme per year (167 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