Explanation-Based Learning for Mobile Robot PerceptionTom M. MitchellJoseph O’Sullivan1Sebastian ThrunSchool of Computer ScienceCarnegie Mellon UniversityAbstractExplanation-based neural network learning (EBNN) has recently been proposed as a methodfor reducing the amount of training data required for reliable generalization, by relying insteadon approximate, previously learned knowledge. We present first experiments applying EBNN tothe problem of learning object recognition for a mobile robot. In these experiments, a mobilerobot traveling down a hallway corridor learns to recognize distant doors based on color cameraimages and sonar sensations. The previously learned knowledge corresponds to a neuralnetwork that recognizes nearby doors, and a second network that predicts the state of the worldafter travelling forward in the corridor. Experimental results show that EBNN is able to use thisapproximate prior knowledge to significantly reduce the number of training examples required tolearn to recognize distant doors. We also present results of experiments in which networkslearned by EBNN (e.g., "there is a door 2 meters ahead") are then used as backgroundknowledge for learning subsequent functions (e.g., "there is a door 3 meters ahead").1. IntroductionOne crucial problem for robot learning is scaling up. Whereas various research has shownhow robots can learn relatively simple strategies from very little initial knowledge (e.g.,[4, 6, 5, 9], theoretical and experimental results indicate that simple learning approaches willrequire unrealistic amounts of training data to learn significantly more complex functions. Oneapproach to scaling up is to rely on human training (e.g., [2, 3, 4, 11]). We are interested here inmethods by which the robot can use previously learned knowledge to reduce the need for newdata in subsequent learning.This paper presents early experiments with a robot architecture that utilizes previously learnedknowledge about the effects of its actions, to reduce the difficulty of learning to recognizeobjects that it approaches as it travels through its environment. This architecture is based onexplanation-based neural network learning (EBNN) [8, 15], an explanation-based algorithm thatrepresents both prior knowledge and the current target function using neural networks. Initialexperiments with EBNN demonstrated its ability to learn robot control knowledge in simulated[8] and real-world robot domains [14]. Here we examine its ability to reduce training datarequired for learning robot perception skills. These results show that EBNN is able tosignificantly reduce the amount of training data required for learning, compared to purelyinductive methods such as Backpropagation, when an appropriate domain theory is available.1Sebastian Thrun is currently at Bonn University.12. Explanation-Based Neural Network LearningExplanation-based neural network learning (EBNN) is a method for using approximate priorknowledge to improve the learner’s ability to generalize correctly from limited training data.Prior knowledge consists of a collection of previously learned neural networks (the domaintheory networks), with the function to be learned represented by an additional neural network(the target network). The EBNN algorithm uses the domain theory networks to analyze eachtraining example, in order to extract information about the relevance of the different features ofthe example. This relevance information is then used, together with standard neural networkinduction, to constrain the weights of the target network. In this way, the target network isconstrained both inductively by the training data, and analytically by the knowledge implicit inthe domain theory.How can the domain theory networks be used to analyze training examples so that informationis extracted in a form useful for refining the target network? The key lies in the TangentPropalgorithm [13], an extension to Backpropagation [12] that is able to adjust network weights tominimize the error in both the values and the derivatives of the function computed by the targetnetwork. Consider some target function, F, and a training example, X, consisting is a vector ofcomponents x . The TangentProp algorithm iteratively adjusts the weights of the target network,iNET, to minimize an error measure containing two terms. The first error term is the differencebetween the target function value, F(X), and the value predicted by the network, NET(X). Thesecond error term is the difference between the partial derivatives of the target function,dF(X)/dx , and the partial derivatives of the function represented by the target network,idNET(X)/dx . EBNN obtains estimates of target values for dF(X)/dx by using its domain theoryi ito analyze training examples.EBNN analyzes a training example, X, by first using the domain theory to predict the value ofF(X). This requires that it be possible to estimate the target function by chaining together thedomain theory networks (as illustrated in Figure 4-1). The weights and activations of the domaintheory networks in this computation are then examined, in order to analytically extract the partialderivative of F(X) with respect to each component feature x . Notice this derivative containsiinformation about the relevance of each feature, x : if the value of x is irrelevant to the value ofi iF it will have a derivative of 0, whereas features with large derivatives will be highly relevant.In this way, the information about feature relevance summarized in the domain theory networksis extracted from the analysis of the example, and used to provide the target derivatives to theTangentProp algorithm, which updates the weights of the target network. In order to minimizethe impact of incorrect domain theories, the degree to which TangentProp seeks to fit these targetderivatives is reduced by the degree to which the domain theory networks err in predicting theobserved trainiing value, F(X).The analytical component of learning in EBNN is similar to earlier explanation-based learningmethods based on symbolic representations [1, 7]: given a target function, F, the domain theoryis used as an alternative method for computing F, and the dependencies are extracted from thiscomputation. The EBNN algorithm is described in greater detail in [8, 15], and summarized herein Figure 2-1.2For each training example:1. Explain how the training example satisfied the target function. Chain together thedomain theory
View Full Document
Unlocking...