The exploratory project InProSSA investigates how different solution paradigms of symbolic and sub-symbolic AI can be combined in a common modeling language. In this way, problems can be formulated independently of the solution approach. The system then automatically selects the best solution method from a pool of existing methods.
Effective decision making in industry requires the optimization of complex problems that need to be adapted to specific goals and constraints. However, difficulties arise when modelling these discrete optimization problems. For example, the question of the appropriate modeling paradigm arises – be it mixed integer programming, constraint programming or heuristic methods. Equally important is the choice of a suitable solver such as Gurobi, IBM CP or Google OR tools.
In addition, challenges arise when business rules change and models need to be adapted. New boundary conditions often influence both the solvability and the runtime of the algorithms. Symbolic solvers often reach their efficiency and scalability limits in industrial applications. Therefore, a flexible approach is needed that decouples the problem description from the solution method and thus allows broader applicability.
InProSSA aims to unite different solver paradigms in a common language for combinatorial optimization problems. The approach of eliminating the boundary between symbolic and sub-symbolic methods is particularly ambitious. Instead, both worlds are to be brought together in a superordinate framework.
Specifically, symbolic solvers such as SAT solvers are used and evaluated. At the same time, the project is investigating modern sub-symbolic methods, such as neural Monte Carlo Tree Search (MCTS). This is based on the assumption that there is no universally best method for all tasks – the so-called no-free-lunch theorem. Each problem therefore requires a specifically adapted solution concept.
InProSSA tests the feasibility of an industrial use case. The focus here is on the integration of existing methods. All results – i.e. articles, software and data – are made publicly available. Based on these findings, new research questions will be developed and implemented in industrial projects in cooperation with company partners.
A local consortium is working together on this challenge. Experts from the fields of formal languages (Research Institute for Symbolic Computation), symbolic AI (Institute for Symbolic Artificial Intelligence), sub-symbolic AI (RISC Software GmbH) and industrial applications (RISC Software GmbH) are involved. The close combination of these competencies creates a novel approach to solving complex industrial tasks more efficiently and flexibly.
This project is funded by the Austrian Research Promotion Agency (FFG).
Mathematical Optimization Specialist
