Automatic apriori decision making for multi-objective optimization

Sammanfattning

Autonomous decision-making in industrial applications is a problem of increasing significance in the age of mathematical optimization and artificial intelligence. In this thesis, a specific decision maker’s preferences are generalized in order to automatically make decisions for any decision problem in a well-defined decision context for multi-objective optimization for food processing. Several criteria need to be optimized simultaneously, and a final alternative needs to be chosen from the resulting pareto-optimal set of alternatives. This optimization and trade-off analysis needs to be possible for many different scenarios where a user can decide what values to target for the criteria. The method proposed in this thesis uses an additive value function, with pre-specified adaptive partial value functions, and a generalized regression function predicting the trade-off weights for any given scenario in the decision-space. Several regressors were evaluated, where a multi-output Random Forest model was deemed superior. The regressor was trained on synthetic swing-weight experiments utilizing Latin Hypercube Sampling, to reduce the number of experiments and ensuring high coverage of the objective-space. The proposed methodology was tested on five simplified quad-objective optimization problems and compared to both manual aposteriori decisions by the decision maker, and an out-of-the-box goal-oriented method using importance weights. The results indicate that the proposed method is better or equally as good as existing methods and can represent the decision maker’s preferences well over the relevant decision space. The proposed method was also tested on five full quad-objective optimization problems, where the pareto-optimal set was significantly larger and infeasible for a decision maker to make a holistic judgement and decision for, proving the need for an automatic decision methodology for similar decision contexts. In conclusion, the proposed methodology shows a promising alternative to existing apriori or interactive methodologies whilst also successfully generalizing the decision-maker’s preferences, enabling automatic decisions in the defined context. This work is important since there are limited studies done in real industrial settings, and for cases using more than two objectives.