Black-Box Optimization Lab. 

The Black-Box Optimization Lab (BBOLab) develops methods for optimizing systems whose inner workings are unknown, using only the outcomes of trying them. When the object to be optimized is a simulation, an experiment, a learning agent, or human judgment, and its internal mechanism or gradient cannot be used directly, the problem is black-box optimization in a broad sense.

Our research proceeds along two parallel pillars. The first is evolutionary computation: centered on CMA-ES, we study its design principles and theoretical analysis, and extend it to high-dimensional and mixed-variable problems. The second is reinforcement learning: we study bilevel reinforcement learning, in which a designer of rewards or incentives and an agent that learns in response are nested; reinforcement learning from human feedback (RLHF) for aligning large language models; and reinforcement learning with safety guarantees. These, too, are black-box optimization problems, since quality can be judged only through another agent’s learning outcome or human judgment. In both pillars, we pursue reliable improvement with few trials, in both theory and implementation.

We publish at the major conferences in evolutionary computation (GECCO, PPSN, FOGA) and in machine learning and AI (NeurIPS, ICML, AAAI, etc.), as well as in international journals. Since 2015, we have given tutorials on CMA-ES at GECCO and other conferences together with Nikolaus Hansen (Inria), the originator of CMA-ES.

We collaborate with researchers at universities, research institutes, and companies in Japan and abroad. If you are interested in joint research or academic consulting, please feel free to contact us.

News