Heteroscedastic Regression and Active Learning for Modeling Affordances in Humanoids

Abstract

We investigate the learning of object and tool affordances in the iCub robot. We adopt the setup proposed in a previous experiment, using a Bayesian network (BN) in place of a least square support vector machine (LSSVM). The collected data, consisting of continuous and discrete variables, are used for learning the structure of the BN. Hence, the model is leveraged to: 1) identify a regression function for the prediction of the effects of actions on objects, calculated as the mean of the observed values; and 2) provide information on the reliability of the predicted values through the estimation of the variance for subsets of local observations. The information on the input-dependent variance is used to guide the learning algorithm in order to improve the performance of the robot, and hence to reduce the variance from the predicted values. The replacement of the LSSVM with the BN model provides a general probabilistic framework for dependencies among random variables; we perform conditional probability queries that enable the robot to choose the actions to perform on objects and select the most appropriate tool to obtain desired effects. The capability to make inference enables the robot to gather a better understanding of the world.