Game-theoretic framework reveals which hyperparameters truly matter
Shapley Effects and Pareto fronts cut hyperparameter search space in multi-objective settings.
A new paper on arXiv proposes a game-theoretic framework for interpretable hyperparameter sensitivity analysis, moving beyond black-box optimization to reveal which hyperparameters actually drive performance. The method applies Shapley Effects—a concept from cooperative game theory—to quantify each hyperparameter's contribution to multiple objectives simultaneously. It then uses Pareto front sets to highlight configurations that dominate across trade-offs, tested on three distinct neural network architectures in multi-objective settings.
This framework does not replace existing optimizers but provides actionable insights: practitioners can shrink the search space by focusing on influential hyperparameters, understand objective-aware interactions, and accelerate early-stage model evaluation. By explaining why certain settings matter, it turns hyperparameter tuning from brute-force grid searching into a guided, interpretable process—particularly valuable for complex models where compute budgets are tight.
- Uses Shapley Effects to assign numerical importance scores to each hyperparameter across different objectives
- Pareto front analysis identifies non-dominated configurations, reducing the search space by focusing on high-impact settings
- Validated on three neural network architectures (e.g., CNNs, RNNs) under multi-objective optimization scenarios
Why It Matters
Practitioners can cut tuning costs and gain clear, objective-specific guidance on which hyperparameters to prioritize.