Popis: |
A learning autonomous robot must learn from sensor data and must decide what topics to learn about. We present the method of resolution limited quantization for learning from sensor data and the method of histogram density to guide the process of topic selection. The methods are complementary in that they use the same knowledge representation. We describe a program, GRID, which implements these methods,. We present an example run of this program learning in the domain of a simulated mobile robot.1. INTRODUCTIONWe address the problem of an autonomous robot using sensors and actuators to learn about its environment. In this context, learning means discovering the relationships among various sensor and actuator values.Elsewhere we describe several issues that arise when machine learning techniques are applied to real autonomous robots1 . Issues addressed in this paper are learning from sensor data and selecting topics.Learning from sensor data is different from the problems addressed by most machine learning research. Machine learning research has typically used non-numeric data or discrete numeric data with only a few possible values for each variable. Methods developed for this limited problem often use exhaustive search of a disjunctive concept space. These methods are computationally intractable in dealing with floating point data, or with very large sets of discrete values (such as inputs from a 16-bit sensor). To reduce the size of the concept space, users can quantize the sensor data, e.g. a temperature range of [1..100] becomes [cold,hot]. Schlimmer^ developed a method that allows a supervised learner to appropriately quantize real-valued data for the task of distinguishing between classes of examples. An autonomous robot cannot use this approach, because the robot is not told how many classes exist nor given clearly labeled examples of each.The problem of learning from sensors is compounded because sensor data are noisy. In the traditional machine learning paradigm, noise means that one or more features of an example are incorrect and should be disregarded. For sensor data, noise usually means that the data are not precisely correct but are close to the actual values. Under this noise paradigm, all examples may be noisy. A robot that disregards all examples will learn nothing, so learning methods which tolerate imprecise data are needed.The issue of topic selection arises because autonomy requires choosing what to investigate and when to terminate a line of investigation, choices previously made by human users. This issue is vital for an autonomous robot, which could spend all its resources on unproductive efforts if it chooses wrong. The issue was addressed by Lenat^ for autonomous learning about mathematics, but his methods are insufficient for more general problems.We present the method of resolution limited quantization as a means to address the problem of learning from sensors. We present the method of histogram density as a means to measure the quality of learned information and therefore to guide topic selection. The methods are complementary to each other and are inexpensive in time and memory. We describe the implementation of these methods in the GRID program, and show the results of applying that program to the domain of a simulated mobile robot with noisy sensors.2. RESOLUTION LIMITED QUANTIZATIONFigure 1 shows five example x,y pairs plotted in x,y space. A robot should be able to generalize from these examples to predict plausible future examples. There are several possible ways to represent such a generalization, e.g. a numeric formula relating x to y or a union of circular regions surrounding each example. Here we will |