Hyperparameter Tuning in Bandit-Based Adaptive Operator Selection

We are using bandit-based adaptive operator selection while autotuning parallel computer programs. The autotuning, which uses evolutionary algorithm-based stochastic sampling, takes place over an extended duration and occurs in situ as programs execute. The environment or context during tuning is either largely static in one scenario or dynamic in another. We rely upon adaptive operator selection to dynamically generate worthy test configurations of the program. In this paper, we study how the choice of hyperparameters, which control the trade-off between exploration and exploitation, affects the effectiveness of adaptive operator selection which in turn affects the performance of the autotuner. We show that while the optimal assignment of hyperparameters varies greatly between different benchmarks, there exists a single assignment, for a context, of hyperparameters that performs well regardless of the program being tuned.