Tree-Structured Parzen Estimator Can Solve Black-Box Combinatorial Optimization More Efficiently

This paper addresses the challenge of applying the Tree-structured Parzen Estimator (TPE) to black-box combinatorial optimization, a task often encountered in fields like chemistry and biology. Traditional TPE struggles with combinatorial problems because it treats all combinations as equally similar. The authors propose an efficient combinatorial algorithm for TPE that incorporates a user-defined distance structure between categories, effectively generalizing the categorical kernel with a numerical kernel. They introduce modifications to handle large combinatorial search spaces and reduce the time complexity of kernel calculations. Experimental results on synthetic problems demonstrate that the proposed method identifies better solutions with fewer evaluations compared to the original TPE. The algorithm is implemented in the Optuna open-source framework. This work bridges the gap between TPE and combinatorial optimization, offering improvements in sample efficiency and performance. #TPE #CombinatorialOptimization #HyperparameterOptimization #BlackBoxOptimization #MachineLearning #Optuna #Algorithm paper - https://arxiv.org/pdf/2507.08053v1 subscribe - https://t.me/arxivdotorg donations: USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e created with NotebookLM