TL;DR
This paper tackles the challenge of optimizing a complex hospital simulation model with a limited budget of expensive function evaluations. They introduce a novel surrogate-based optimization approach using a random ReLU expansion, demonstrating its superior performance over established methods like Bayesian Optimization and Hyperopt for this specific industrial problem.
What Problem Does This Paper Address?
Optimizing complex simulation models, especially those used for critical applications like hospital resource planning, is a significant challenge. These simulations often have a very long run time, making it impractical to evaluate the objective function many times. The GECCO 2021 industrial challenge presented exactly this scenario: tuning 29 parameters of a hospital simulator (BaBSim.Hospital) to achieve optimal resource allocation, particularly relevant during the COVID-19 pandemic. The simulator's long run time (average 86 seconds per evaluation) and non-deterministic nature, coupled with a strict limit of only 200 function evaluations, necessitate highly efficient optimization strategies. The goal is to find the parameter set that minimizes the objective score, subject to defined bounds for each parameter.
Methods
The authors employ a surrogate-based optimization framework. This approach involves building a cheaper-to-evaluate approximation (a surrogate model) of the expensive objective function. The optimization process iteratively:
- Evaluates the true objective function at a candidate point.
- Updates the surrogate model with this new data point.
- Uses the surrogate model to suggest the next promising point to evaluate.
The core novelty of their work lies in the surrogate model itself: a random ReLU expansion (RRE). This model is a type of random feature expansion that uses rectified linear units (ReLUs) as basis functions. The RRE model is defined as $g(x) = \sum_{k=1}^D c_k \phi(w_k^T x + b_k)$, where $\phi$ is the ReLU function, and $w_k$, $b_k$ are randomly chosen weights and biases that remain fixed. The coefficients $c_k$ are then trained using a recursive least squares algorithm. This RRE model is piece-wise linear.
To guide the search, they use a continuous optimizer, L-BFGS, to find a local minimum of the surrogate model $g(x)$. Crucially, they then add a random perturbation to this point to encourage exploration and avoid getting stuck in local optima. This perturbation is drawn from a normal distribution with a standard deviation scaled by the parameter bounds and the number of dimensions.
Datasets
The paper evaluates their approach on the GECCO 2021 industrial challenge problem, which involves optimizing the BaBSim.Hospital simulation model. They do not introduce a new dataset but rather use an existing benchmark suite for expensive optimization problems, EXPObench [7], to compare their method against others.
Key Insights and Ideas
- RREs as effective surrogates for expensive simulations: The random ReLU expansion proved to be a powerful surrogate model for this complex, noisy, and expensive-to-evaluate hospital simulation problem.
- The trade-off between exploration and exploitation: The initial number of random evaluations (r) significantly impacts performance. A larger 'r' provides more data for the surrogate but leaves fewer evaluations for the guided search.
- Random perturbation is key for exploration: Adding a random perturbation to the surrogate's suggested point is crucial for escaping local optima and exploring the search space more effectively.
- Simplicity can be powerful: Unlike some other RRE approaches, this work does not enforce convexity or integer constraints, suggesting that a simpler RRE can be highly effective.
Key Results
The RRE approach, when configured with 58 initial random samples ($r=2d$), achieved the lowest average best objective value of 16.29 across five runs, outperforming Random Search (17.65), HyperOpt (16.81), and Bayesian Optimization (16.78).
However, this configuration also exhibited high variance (±2.16). By halving the number of initial random samples to 29 ($r=d$), the RRE approach achieved an even better average best objective value of 14.81 with significantly reduced variance (±0.69). This highlights the critical balance between initial exploration and subsequent exploitation.
Limitations and Future Work
The authors acknowledge that the initial configuration of their RRE approach (with $r=58$) suffered from high variability. While halving the initial random samples improved both performance and reduced variance, they suggest further work is needed. Future research directions include investigating other parts of their approach, such as the exploration strategy, to achieve even better results on this challenging problem.
Summary updated