Towards Extracting Adaptation Rules From Neural Networks

Defining adaptation rules is an important step in the design of adaptive systems. This paper proposes using a constrained multi-modal neural network to extract adaptation rules. The proposed approach enhances a serious game’s adaptive capability, which aims to help learners improve their socio-moral reasoning skills. The neural network takes learners’ multimodal data as input and predicts how it will answer an exercise. The rules extraction is based on reading and interpreting weights learned by the trained network to determine the players’ attributes and the system’s elements that play an important role in predicting the knowledge involved. The extracted rules are then validated using a decision tree. This validation shows that the proposed technique can support the production of adaptation rules in adaptive systems. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.