Due to demographic change, the number of seniors in nursing homes will increase over the next few years. Since caring for people with dementia is very time-consuming and nursing homes are cutting staff due to cost pressures, the care sector faces enormous challenges in the coming years. However, regular training of certain physical exercises adapted to the individual state of health is very important for the health promotion and resilience of people with dementia. Therefore, it is necessary to explore a dementia robot that enables seniors to perform these exercises independently and flexibly. As dementia exercises that can be transferred to a robot, ball throwing, high-five game and strength exercises have already been derived in preliminary work. Based on this, an adaptive, optimizing and real-time interaction system has now been researched, which uses rule-based Fuzzy Logic to classify the degree of dementia and Evolutionary Algorithms to adapt the exercise parameters to the state of health. To develop the classifier, the expert knowledge of caregivers was collected using knowledge acquisition. The expert knowledge was formalized in a knowledge base using a hybrid inference mechanism. The results show that the degree of dementia can be correctly classified for the ball throwing, high-five game and strength exercise. Similarly, the exercise parameters can also be optimized with respect to individual therapy progress. The interaction system was successfully tested in a real-time robot simulation. Initial tests with a real two-arm robot were successfully performed. The evaluation with trained personnel as well as the acceptance study in a nursing home are in preparation.