Learning efficient logical robot strategies involving composable objects

Conference item

Most logic-based machine learning algorithms rely on an Occamist bias where textual complexity of hypotheses is minimised. Within Inductive Logic Programming (ILP), this approach fails to distinguish between the efficiencies of hypothesised programs, such as quick sort (O(n log n)) and bubble sort (O(n2)). This paper addresses this issue by considering techniques to minimise both the textual complexity and resource complexity of hypothesised robot strategies. We develop a general framework for the problem of minimising resource complexity and show that on two robot strategy problems, 1) Postman 2) Sorter (recursively sort letters for delivery), the theoretical resource complexities of optimal strategies vary depending on whether objects can be composed within a strategy. The approach considered is an extension of Meta-Interpretive Learning (MIL), a recently developed paradigm in ILP which supports predicate invention and the learning of recursive logic programs. We introduce a new MIL implementation, MetagolO, and prove its convergence, with increasing numbers of randomly chosen examples to optimal strategies of this kind. Our experiments show that MetagolO learns theoretically optimal robot sorting strategies, which is in agreement with the theoretical predictions showing a clear divergence in resource requirements as the number of objects grows. To the authors’ knowledge this paper is the first demonstration of a learning algorithm able to learn optimal resource complexity robot strategies and algorithms for sorting lists.

Authors: Cropper, A; Muggleton, S

• Publication status: Published
• Peer review status: Reviewed (other)
• Publisher: AAAI Press / International Joint Conferences on Artificial Intelligence
• Host title: Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence
• Journal: IJCAI
• Issue: 3423
• Pages: 3423-3429
• Publication date: 2015-07-31
• ISBN: 9781577357384
• Keywords: SBTMR