Reproduce “Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Melis Ilayda Bal, Pier Giuseppe Sessa, Mojmir Mutny, Andreas Krause
GameOpt searches sequences by coordinating amino-acid choices at each position around a shared optimistic fitness estimate, rather than scoring every complete sequence. It yields a stable surrogate-search point, not necessarily a global optimum. On GB1(4), GameOpt-IBR found the global best sequence in 33.33% of 18 runs, versus 0% for GP-UCB; this does not establish global success on larger landscapes.
Key insights
- Decomposing the complex, intractable combinatorial acquisition function maximization into a cooperative game among individual variables provides a scalable and tractable optimization strategy.
- Game equilibria (e.g., Nash equilibria) of an optimistic acquisition function can effectively guide the search in unstructured discrete spaces, serving as analogous local optima.
- The framework provides theoretical sample-complexity guarantees, quantifying the number of iterations required to achieve approximate equilibria of the true objective function.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.