Forecasting continuous optimization algorithm performance improvement within each run

This research introduces predictive models to forecast continuous optimization algorithm performance improvement during runtime, departing from conventional approaches focusing solely on final performance outcomes. These models leverage sequential, iteration-scaled data, which includes both the solution locations and their corresponding objective values observed throughout the algorithm’s execution. Using this information, the models aim to forecast performance improvements in future iterations, evaluated over a predetermined horizon points. These real-time forecasting capabilities could be utilized in algorithm configuration scenarios to provide additional insights, further enhancing algorithm performance in black-box continuous optimization problems. We evaluate several long short-term memory networks, selected for their effectiveness with sequential data, in two scenarios: development of personalized models using runs from a particular algorithm instance on a single problem instance and development models using runs from a particular algorithm instance on diverse set of instances of optimization problems. Employing the CEC2014 benchmark suite with 10-dimensional unconstrained problem instances and various configurations of Differential Evolution algorithm, our approach demonstrates promising results by exceeding a baseline forecasting model. • Within-run models predict future performance gains per iteration of problem instance. • Uses scaled population data, ensuring invariance to simple transformations. • Incorporates sequential population dynamics to forecast performance improvements. • Uses LSTM time-series models to capture trends in scaled optimization data. • LSTM consistently outperforms baseline models in forecasting performance gains.