How Detailed Should be a Prediction for a Behaviour Success: Precision of Prediction Information Gain Trade-O in Predictive Processing

Abstract

Recent studies in Predictive Processing use categorical variables for modelling a generative model. This leads yet unexplored consequences of the detail of the prediction de ned by the number of outcomes of the distribution. Low level of detail of the prediction gives more precision of the prediction while giving less information gain on how to the achieve our goal, conversely for a high level of detail. The aim of this project is to implement a model that can select the right level of detail of the prediction to achieve its goal, while making precise predictions and how this could be related to human cognition. In a successful implementation the following hypotheses should hold: rst, an agent using the model will perform a given task successfully, second, the level of detail of the prediction will converge to a single value after enough iterations performing the task, and lastly, on hard tasks, the model will converge to a higher level of detail than on easy tasks. The model controls a robot with the task of going to an object in distance. The performance of the robot is poor, losing the object out of sight quite often and never achieving its goal. The main reasons for the failure are investigated, which include the brittle rules designed to modify the detail of the prediction, overly simplistic model and an inappropriate resolution of the exploration/exportation dilemma. As for modelling human cognition, the need for a global optimisation of the precision and information gain with respect to the detail of the variables in the generative model is discussed.