Abstract:
In control applications, controllers for different plants are usually designed with different methods. Although these plants share common characteristics, these are generally designed in isolation. Recently , several researchers have studied the problem of continuously learning a sequence of related learning tasks. A challenge in continual learning is the phenomenon of catastrophic forgetting of knowledge of previous tasks which have been integrated into a neural network model. In this paper we evaluate the feasibility of modelling different controllers using continual learning. We explore regression versions of state-of-the-art methods and demonstrate that even the simplest continual learning approach decreases the overall Mean Average Error (MAE) by 39% of the MAE achieved by a non-continual strategy. Furthermore, a method based on dynamically expanding the network can achieve an overall MAE which is only 18% of the non-continual MAE. Given these results, we also propose a set of new metrics that allow us to characterise the nature of catastrophic forgetting that occurs for these continual learning methods.