Lecture | Seminar
CCLS Seminar: Giovanni Misitano
- Monday 14 November 2022
Explainable AI and explainable multiobjective optimization
Multi-objective optimization handles problems with multiple conflicting objectives with various tradeoffs: low cost leads to a high, negative environmental impact, for instance. Multi-objective optimization problems have many compromise solutions, which are not comparable on a pure mathematical basis. The aid of a domain expert, known as the decision maker, is employed. The decision maker can provide preference information, which can then be used to find the best possible solution among the compromise solutions.
However, as decision-support tools, multi-objective optimization methods are often used to make real-life decisions with real-life consequences. If we do not fully understand how the tools we use operate, can we really justify the decision we make using them? I argue that many of the methods utilized in multi-objective optimization are pretty much just black-boxes to a decision maker. Can we address this problem?
Explainability has now been an established concept in the research of AI. It has spawned its own research field: explainable AI. Explainability has been used to address many of the problem encountered in AI, such as identifying biases in the data used and to better understand predictions made by black-box models. As a concept, however, explainability can be expanded to address any decision-support tools---not just AI.
I have been pioneering a new paradigm in decision-support tools: explainable multi-objective optimization. In my past works, I have taken inspiration from explainable AI, but in the future, we must developed novel and unique ways of incorporating explainability in multi-objective optimization to address its specific needs. As we move more and more towards a society governed by data and data-based solutions, it is important that we understand the tools we use and that the tools themselves are able to explain their behavior. Because only the best decisions can be also justified.
The CCLS Seminar is not publicly accessible in real-time. Please click the the link below to register If you would like to join.Click here to register!