Regularised Loss Function for Goal Recognition as a Deep Learning Task

Goal Recognition (GR) consists of recognising the goal of an agent from partial observations. The state of the art on particular planning domains is represented by GRNet, a model based on Recurrent Neural Networks that solves GR as a classification task. Compared to automated planning, the need for large training sets is the main disadvantage of these approaches. Therefore, we formalise a loss regularisation technique to reduce the number of training samples needed, to reduce the convergence time, and to increase the performance in GR instances with a small percentage of observations. We empirically evaluate its effectiveness through extensive experiments.