Zimin Liang

Nature harmony, well-being and personalised education for all.

In one line: I study how a multi-objective optimiser should keep, discard, revisit, and generate solutions when the search space is hard — the setting behind everything from factory scheduling to assembling an LLM agent team — and I find that the simple choice is often the right one.

Most multi-objective evolutionary algorithms are built on a sensible instinct: protect the best solutions you have found so far. That instinct is usually right. But it is not the whole story. On a rugged or deceptive landscape, strong preservation can trap a search — it keeps exploiting a region that looks good locally but leads nowhere.

This line of work asks how to keep the useful parts of elitism and archiving while leaving a search enough freedom to escape such traps.

The design question

In a multi-objective problem, an algorithm maintains a population or an archive of trade-off solutions. Every generation forces one question:

Which solutions should be allowed to influence the future of the search?

It shows up in several concrete forms:

  • Should the old population always survive if it is currently good?
  • Should weaker solutions sometimes survive, because they may open the door to better regions later?
  • Should an archive only record solutions, or actively generate new search steps from them?
  • In local search, should a neighbourhood be scanned in order, or sampled at random?

I approach these through both experiments and formal proof.

Non-elitist MOEAs

Elitism — selecting the next population from both old and new solutions, so good ones are never lost — has been a near-universal default since the late 1990s. In GECCO 2023, we asked what happens if you deliberately remove it.

The resulting non-elitist MOEA discards the old population entirely and lets selection pressure come from mating selection alone. The point was not that elitism is bad, but that the design space is wider than the standard template admits. On several rugged combinatorial problems, this simple non-elitist idea held its own against established elitist algorithms such as NSGA-II, SMS-EMOA, and NSGA-III.

The takeaway: preservation and progress are not the same thing. A population can hold high-quality solutions while being badly positioned for the next move.

Stochastic update with archives

Stochastic population update lets some weaker solutions survive at random. This can help an algorithm jump out of a local optimum — but it also risks losing genuinely good solutions.

In IJCAI 2025, we studied this tension formally. The result explains why an archive matters when update is stochastic: the archive protects the best solutions found so far, while the stochastic population provides room to explore. With this pairing, a small population can be enough, and the expected running time can improve sharply.

The principle is simple: randomness is useful, but it should be paired with memory.

Archive reuse

An archive is usually treated as a passive record of the non-dominated solutions seen so far. In AAAI 2026, we gave it an active role: feeding archived solutions back into offspring generation.

The motivation is intuitive. A small population may have to delete a promising solution simply because it cannot hold everything. If the archive only stores that solution, the algorithm remembers it but never uses it again. If the archive can be reused, the search can return to that promising region.

The analysis shows this reuse can deliver polynomial speed-ups on benchmark problems — and can even beat the alternative of simply using a larger population.

Local search feels most efficient when it scans a neighbourhood systematically: inspect each neighbour once, never revisit, stop on a condition. In multi-objective local search, that intuition can fail.

In AAAI 2026, we showed that random neighbourhood sampling can be faster than systematic exploration across a range of multi-objective problems. The reason is the distribution of “good neighbours” during the search: when useful moves are spread in a way that rewards repeated random trials, the bookkeeping overhead of systematic scanning outweighs its apparent tidiness.

This is a good example of the style of work I care about: take a familiar assumption, find exactly where it breaks, then explain the mechanism behind the result.

Decoupling exploration and exploitation

The sections above keep returning to one tension: a population must both exploit the good solutions it has and explore toward new regions, and the solutions best suited to one job are often the worst for the other. The usual response is to tune a single population to compromise between the two.

A cleaner answer is to stop compromising. A dual-population design assigns each role its own population: one keeps only the best (non-dominated) solutions for exploitation, while the other ignores quality entirely and is governed by a simple aging rule, so that recently generated solutions — including poor ones with exploratory value — survive long enough to be useful. A single probability decides which population a parent is drawn from, which makes the intensity of exploration something you set directly rather than something you hope emerges from one population’s dynamics. This decoupling can yield large, provably faster search on hard multimodal problems.

Where the design work goes next

The minimal-algorithm thread continues to open questions: how to extend a simple optimiser such as SEMO to variable-length representations, and how to make its search steps adaptive rather than fixed. These are ongoing and will appear here as the work matures.

Selected publications

For the full published record, see the Publication page.

  • Z. Liang, M. Li. Random is Faster than Systematic in Multi-Objective Local Search. AAAI, 2026. [PDF] [Code]
  • S. Ren, Z. Liang, M. Li, C. Qian. Not Just for Archiving: Provable Benefits of Reusing the Archive in Evolutionary Multi-objective Optimization. AAAI, 2026. [PDF] [Code]
  • S. Ren, Z. Liang, M. Li, C. Qian. A Theoretical Perspective on Why Stochastic Population Update Needs an Archive in Evolutionary Multi-objective Optimization. IJCAI, 2025. [PDF]
  • Z. Liang, M. Li, P. K. Lehre. Non-elitist Evolutionary Multi-objective Optimisation: Proof-of-principle Results. GECCO Companion, 2023. [PDF] [DOI]