Applying neuroevolution to estimate the difficulty of learning activities

Abstract

Learning practical abilities through exercises is a key aspect of any educational environment. To optimize learning, exercise difficulty should match abilities of the learner so that the exercises are neither so easy to bore learners nor so difficult to discourage them. The process of assigning a level of difficulty to an exercise is traditionally manual, so it is subject to teachers’ bias. Our hypothesis is about the possibility of establishing a relation between human and machine learning. In other words, we wonder if exercises that are difficult to be solved by a person are also difficult to be solved by the computer, and vice versa. To try to bring some light to this problem we have used a game for learning Computational Logic, to build neuroevolutionary algorithms to estimate exercise difficulty at the moment of exercise creation, without previous user data. The method is based on measuring the computational cost that neuroevolutionary algorithms take to find a solution and establishing similarities with previously gathered information from learners. Results show that there is a high degree of similarity between learner difficulty to solve different exercises and neuroevolutionary algorithms performance, suggesting that the approach is valid.