Design of intelligent mechatronical systems with modifiable behaviors

We present and extend an approach for the integration of reinforcement learning methods into Petri net based specifications of autonomous behaviors. The work aims at the design of autonomous mechatronical systems with modifiable adaptive behavior and our extension handles the required modifiability. In order to combine Petri nets and learning methods, we modeled Q-learning - a variant of reinforcement learning - with high-level Petri nets. The result can be integrated into Petri net models of autonomous mechatronical systems. For an evaluation of our approach, we have implemented a realistic application example. It has been evaluated by simulation as well as on a physical system