Modular Learning and Optimization for Planning of Discrete Event Systems

Doctoral thesis, 2021

Optimization of industrial processes, such as manufacturing cells, can have great impact on their performance. Finding optimal solutions to these large-scale systems is, however, a complex problem. They typically include multiple subsystems, and the search space generally grows exponentially with each subsystem. This is usually referred to as the state explosion problem and is a well-known problem within the control and optimization of automation systems.

This thesis proposes two main contributions to improve and to simplify the optimization of these systems. The first is a new method of solving these optimization problems using a compositional optimization approach. This integrates optimization with techniques from compositional supervisory control using modular formal models, dividing the optimization of subsystems into separate subproblems. The second is a modular learning approach that alleviates the need for prior knowledge of the systems and system experts when applying compositional optimization.

The key to both techniques is the division of the large system into smaller subsystems and the identification of local behavior in these subsystems, i.e. behavior that is independent of all other subsystems. It is proven in this thesis that this local behavior can be partially optimized individually without affecting the global optimal solution. This is used to reduce the state space in each subsystem, and to construct the global optimal solution compositionally.

The thesis also shows that the proposed techniques can be integrated to compute global optimal solutions to large-scale optimization problems, too big to solve based on traditional monolithic models.