/ research programme
Generative Optimization
Generating designs, judging them by how people experience them, and searching for better ones.
The organising idea behind my work since 2012. Generative AI, games informatics and metaheuristics are not three separate interests: they are the three stages of one loop. Where it goes next is Generative Optinformatics.
/ working definition
Generative Optimization treats creative design as a search problem: generative models propose candidate artefacts, models of human engagement evaluate them, and metaheuristic search steers what is generated next.
The goal is not a single optimum but a set of artefacts that are good, diverse and genuinely engaging for the people who use them. What sets the programme apart is the middle stage: the objective comes from a theory of engagement developed and published over more than a decade, rather than from a generic quality metric.
Illustrative teaching model: random edits stand in for the AI generator, and GR = √G / T is a simplified game-refinement measure.
/ one loop, three stages
Each research line is one stage.
The three pillars on this site map directly onto the loop. Generation without evaluation produces volume, not quality. Evaluation without search only ranks what already exists. Search without a human-centred objective optimizes the wrong thing. The contribution is in closing the loop.
What could exist?
from: Generative & Agentic AI
Generative models, LLMs and agent teams propose candidate artefacts: levels, quests, puzzles, layouts, schedules, even new search operators. Generation is cheap and plentiful; on its own it has no sense of what is worth keeping.
Is it any good, for people?
from: Games Informatics & Engagement Modelling
Each candidate is scored by models of how people experience it. Game-refinement theory and the "motion in mind" model turn engagement into a measurable signal, so the objective is human experience, not only accuracy or cost.
How do we find better ones?
from: Computational Intelligence & Optimization
Metaheuristics decide what gets generated next: evolutionary, immune and swarm algorithms, and quality-diversity archives such as MAP-Elites that keep many different good solutions rather than one winner.
/ how it relates
Built on established fields.
Generative Optimization draws on several active research communities. It shares their methods and is precise about what it adds.
/ research horizon
From optimizing designs to learning how to optimize.
Generative Optimization
Close the generate → evaluate → search loop for individual design problems: puzzles, levels, quests, learning activities and interactive systems, each optimized against human engagement.
Generative Optinformatics
Make the loop learn from itself. Every run leaves knowledge behind: archives of elites, search trajectories, evaluator feedback. Generative models can mine that experience and reuse it, to seed new searches, propose better operators and transfer what worked from one design problem to the next.
Builds on optinformatics, the established idea of bringing informatics and learning into the optimization process (Lim et al. 2016; Feng, Hou & Zhu 2021), extended here with generative models.
Key references
- Lim, D., Ong, Y.-S., Gupta, A., Goh, C. K. & Dutta, P. S. “Towards a new Praxis in optinformatics targeting knowledge re-use in evolutionary computation: simultaneous problem learning and optimization.” Evolutionary Intelligence 9(4): 203–220 (2016).
- Feng, L., Hou, Y. & Zhu, Z. Optinformatics in Evolutionary Learning and Optimization. Adaptation, Learning, and Optimization, vol. 25. Springer (2021). Link
- Togelius, J., Yannakakis, G. N., Stanley, K. O. & Browne, C. “Search-based procedural content generation: A taxonomy and survey.” IEEE Transactions on Computational Intelligence and AI in Games 3(3): 172–186 (2011).
- Mouret, J.-B. & Clune, J. “Illuminating search spaces by mapping elites.” arXiv:1504.04909 (2015). Link
- Liu, S., Chen, C., Qu, X., Tang, K. & Ong, Y.-S. “Large language models as evolutionary optimizers.” arXiv:2310.19046 (2023, rev. 2024). Link
My own work across all three stages is on the publications page.
Work on the loop with me.
PhD and Master's projects are open at every stage: new generators, new engagement measures, and new search methods. Industry partners with a design problem worth optimizing are welcome too.