The goal in dynamic optimization is to track the moving optima of time-varying fitness functions. Successful optimization algorithms are evolution strategies treating the fitness function as a black-box. Evolution strategies can be equipped with a prediction model that forecasts the new optimum after a fitness function change. The predicted optimum can then be utilized during optimization in order to find the true optimum faster and more accurately. In contrast, applying a prediction model can also mislead the optimization in case the predicted optimum differs much from the real one, i.e., the predictive uncertainty is large.

In my thesis, I propose to use artificial neural networks as prediction models in evolutionary optimization. Further more, I extend two evolutionary heuristics, i.e., evolution strategy and particle swarm optimization, with a prediction approach. In addition, I estimate the predictive uncertainty and propose a way to consider it during optimization.

According to the experimental results on different benchmark problems, neural network predictors can capture more complex problem dynamics than conventional approaches like autoregression and Kalman filters. Considering predictive uncertainty estimation during optimization supports evolution strategies to find good solutions directly after a fitness function change. In the remaining PhD project time, I plan to investigate prediction for multimodal problems, i.e., problems with many local optima.