Part 9/14:
Extending Bayesian Optimization to Parallel Computing
A major limitation of classical BO is its inherently sequential nature: each new point depends on the previous evaluation. To accelerate exploration, Shell’s team experimented with parallel Bayesian Optimization, where multiple evaluations are performed simultaneously.
This approach involved:
Selecting multiple promising points (e.g., four at a time) based on an acquisition function adapted for batch evaluation.
Updating the surrogate after all parallel evaluations are completed.
Significantly reducing total optimization time—achieving up to 4x speedup over traditional sequential methods.
Results and Insights
Implementing parallel Bayesian Optimization led to notable efficiency gains, including: