Current Research Directions
Our main research interests include:
- Interactive multiobjective optimization methods
- Real-world applications
- Data-driven decision support, decision (prescriptive) analytics and simulation-based problems
- Open source software development
- Explainable artificial intelligence in decision support
- Deep uncertainty
- Group decision making
- Visualizations
Examples of thesis topics relating to the current research
direction for bachelor, master and doctoral students are
available here.
Examples of the multiobjective optimization methods we have developed:
- NIMBUS - classification-based interactive method and its variants for robust and scenario-based problems
- Pareto Navigator and Nonconvex Pareto Navigator – interactive real-time navigation methods for computationally expensive problems
- NAUTILUS, NAUTILUS-2, E-NAUTILUS, NAUTILUS Navigator, O-NAUTILUS Navigators - family of interactive methods where the decision maker can gain without sacrifices
- PAINT - method preparing a computationally inexpensive surrogate problem
- K-RVEA - evolutionary method for computationally expensive problems
- interactive RVEA – interactive evolutionary method with four types of preference
information
- PBEA – interactive version of IBEA
- Interactive K-RVEA – interactive evolutionary method for computationally expensive problems
- SURROGATE-ASF – interactive surrogate-based method for computationally expensive problems
- IOPIS - a new paradigm that makes any evolutionary method interactive
The most well-known interactive method developed in our group is NIMBUS. NIMBUS is a classification-based method where a decision maker classifies objective functions to indicate the kind of changes that are desired in the current Pareto optimal solution to make it better. Several variants of the method have been published during the years and the synchronous version is currently in use.
Among more recently developed interactive methods we can mention Pareto Navigator and Nonconvex Pareto Navigator which have been directed for computationally expensive problems. The idea is to create an approximation of the Pareto optimal set and enable the decision maker to navigate on it. On the approximation, changes of trade-offs can be seen in real-time and then any interesting solution can be projected to the real Pareto optimal set. Without the approximation, the navigation would be too slow because calculating new Pareto optimal solutions would take too much time.
On the other hand, the family of interactive NAUTILUS methods questions the idea of considering only Pareto optimal solutions throughout the solution process. Instead, the method starts from an inferior point like the nadir point and allows finding the most preferred solution without anchoring and the need of giving up in any objective functions. This can be useful also for group decision making situations. We have developed several variants of NAUTILUS where the decision maker has different ways to specify preference information to direct the solution process
Because many methods developed in our group are motivated by practical applications, it is important that this work is also brought close to people that in real-life face the actual problems and apply the methods. Therefore, our group has a history of preparing implementations of the methods developed.
One can say that the first widely available interactive multiobjective optimization software has been WWW-NIMBUS (available at
http://nimbus.it.jyu.fi/), an implementation of the NIMBUS method. WWW-NIMBUS is a web-based software freely available for academic teaching and research use around the world (the first version was published as early as in 1995). Based on the WWW-NIMBUS software, the group also developed IND-NIMBUS, which is a desktop application operating on Linux and Windows platforms. IND-NIMBUS has been used for various industrial applications. A demo-version is available for interested parties (http://ind-nimbus.it.jyu.fi/).
For several years, the our group has worked on developing the open source software framework DESDEO where interactive methods developed by us and others can be found. In the implementations, we pay special attention to intuitive human-computer interaction and visualizations. Because interactive methods are supposed to support learning, the way preference information is acquired from and new insight into the problem is presented to the decision maker plays a very important role in the success of the solution process. This research contains user interface and interaction design, usability research, information visualization and visual analytic environments.
Our group has been active in building bridges between the multiple criteria decision making (MCDN) and evolutionary multiobjective optimization (EMO) communities. Examples of hybrid method development include approaches for estimating the nadir point utilizing EMO and achievement scalarizing functions. In addition, the efficiency and accuracy of EMO methods have been improved by hybridizing scalarizing functions and local search in them. Preference information has been also incorporated in EMO methods in the form of e.g. a reference point improving efficiency and enabling concentration on interesting Pareto optimal solutions.
Alongside the general theoretical and methodological development, the development of so-called approximation methods has been considered. These methods utilize an approximation of the objective functions or the Pareto optimal set (e.g. metamodels, polyhedral and tangent plane approximations). An approximation of the Pareto optimal set is especially useful in the case of industrial applications because problem related models are typically computationally very time-consuming to operate.
Another research direction is related how to tackle with uncertainty in multiobjective optimization problems, that is, how to compare solution alternatives under uncertainty and in changing environments. We have also extended our focus from supporting a single decision maker to groups of decision makers.
