Selbstverstärkendes Lernen als Beitrag zur Automatisierung der Anpassungskonstruktion

Description

Automation describes the transfer of tasks from humans to machines. In the context of adaptation design, these tasks involve defining the product´s characteristics to meet new requirements while maintaining the same principal solution. They are particularly challenging if characteristics simultaneously influence competing properties. Here, the best possible compromise must be found. Depending on the context of the adaptation task, they can be automated by optimisation algorithms or by data-driven approaches. However, data-driven approaches require that the underlying database contains all relevant information for recognising the relationships between characteristics and resulting properties. For their part, optimisation algorithms must be brought to convergence whenever the boundary conditions change. For the intersection of adaptation tasks, for which the boundary conditions change regularly due to changes in requirements and at the same time no sufficient database is available, this work presents reinforcement learning as an extension of the current automation possibilities. Reinforcement learning is described as the learning of an agent based on experience with an environment. Here, the environment represents the tasks of the product developers in the context of adaptation design. In which, the agent is trained to take over tasks by adapting characteristics and receiving the product properties as feedback in return. This work does not to develop a new algorithm for reinforcement learning, but takes reinforcement learning into product development, defines the role of the product developers in its use and to research the transferability of the experience gained from one adaptation task to a subsequent one.