Research Unit for Intelligent Optimization Technologies
Optimization Research Base for Intelligent Technologies
From a person-dependent craft to a reproducible scientific foundation for deep-learning hyperparameter search.
We fuse evolutionary computation with combinatorial optimization and introduce a new metric —
Optimization Ease Score (OES) — that captures accuracy, stability, reproducibility and efficiency.
ORBIT (Optimization Research Base for Intelligent Technologies) is dedicated to transforming the complex parameter search that governs deep-learning performance — today reliant on experience and intuition — into a reproducible and sustainable technological foundation. By integrating evolutionary computation and combinatorial optimization, ORBIT develops a search framework capable of stably achieving high performance under limited computational resources, even for realistic problems with both continuous and discrete variables. We propose a new metric, the Optimization Ease Score (OES), that evaluates not only performance but also attainment, stability and reproducibility.
Advanced PSO and evolutionary algorithms reach high-quality solutions efficiently — even in noisy, multimodal real-world landscapes.
Local search, neighborhood design, and constrained exploration efficiently navigate model structures and discrete configurations.
Beyond peak accuracy: we quantify how stably, reproducibly and cheaply anyone can reach high performance, making AI development truly reproducible.
Today's deep-learning workflow depends on expert trial-and-error, sacrificing both cost and reproducibility. ORBIT combines two optimization technologies with a new evaluation metric, OES, to deliver a next-generation AI development platform anyone can use.
Substantial computational resources are wasted on tuning, inflating development cost.
Reliance on expert experience reduces reproducibility and concentrates expertise in individuals.
Continuous and discrete variables coexist — most existing tools cannot handle them jointly.
Swarm and evolutionary methods efficiently search continuous hyperparameters.
Heuristic search and neighborhood design unify mixed-variable HPO.
A metric that includes stability, reproducibility and efficiency, not just accuracy.
A reliable, accessible foundation enabling anyone to deliver high-performing AI.
Early stopping, weight inheritance and parallelization reduce cost and footprint.
Results are released openly to advance academia and industry.
ORBIT proposes Optimization Ease Score (OES) as a new axis for evaluating deep-learning models. Conventional benchmarks compare peak accuracy alone, ignoring the questions that matter in practice: does it work for anyone, and can it be reproduced under realistic budgets?
OES evaluates a model together with its optimization procedure across four dimensions — attainment, stability, reproducibility, and computational efficiency — capturing how usable and how shareable an AI system really is.
Background, challenges, approach and vision — at a single glance.
Four researchers spanning deep learning, evolutionary computation, combinatorial optimization, and parallel computing form ORBIT.
Professor — Deep Learning & OES Design
Building on deep learning and internal-representation analysis, leads task and evaluation design, the definition and validation of OES, integration of the framework, and overall program management — expanding model evaluation from accuracy to stability, reproducibility and robustness.
Professor — Evolutionary Computation
Leads continuous-variable hyperparameter optimization design, leveraging research in evolutionary computation and swarm intelligence — advancing PSO and noise-robust search to support deep-learning HPO.
Professor — Combinatorial Optimization
Drives the design of discrete and structural-variable search, applying local search, neighborhood design and constrained heuristics to model-structure selection and combinatorial configuration of deep-learning models.
Associate Professor — Implementation & Acceleration
Responsible for parallelization, high-performance implementation and computational-resource efficiency — enabling large-scale searches to run on practical hardware budgets.