Most artificial intelligence systems today work on simple problems and artificial domains because they rely on the accurate sensing of the task world. Object recognition is a crucial part of the sensing challenge and machine learning stands in a position to catapult object recognition into real world domains. Given that, to date, machine learning has not delivered general object recognition, we propose a different point of attack: the learning architectures themselves. We have developed a method for directly learning and combining algorithms in a new way that imposes little burden on or bias from the humans involved. This learning architecture, PADO, and the new results it brings to the problem of natural image object recognition is the focus of this chapter.

This thesis investigates the evolution and use of abstract data types within Genetic Programming (GP). In genetic programming the principles of natural evolution (fitness based selection and recombination) acts on program code to automatically generate computer programs. The research in this thesis is motivated by the observation from software engineering that data abstraction (e.g. via abstract data types) is essential in programs created by human programmers. We investigate whether abstract data types can be similarly beneficial to the automatic production of programs using GP. GP can automatically "evolve" programs which solve non-trivial problems but few experiments have been reported where the evolved programs explicitly manipulate memory and yet memory is an essential component of most computer programs. So far work on evolving programs that explicitly use memory has principally used either problem specific memory models or a simple indexed memory model consisting of a single glo...

Most artificial intelligence systems today work on simple problems and artificial domains because they rely on the accurate sensing of the task world. Object recognition is a crucial part of the sensing challenge and machine learning stands in a position to catapult object recognition into real world domains. Given that, to date, machine learning has not delivered general object recognition, we propose a different point of attack: the learning architectures themselves. We have developed a method for directly learning and combining algorithms in a new way that imposes little burden on or bias from the humans involved. This learning architecture, PADO, and the new results it brings to the problem of natural image object recognition is the focus of this report. 1 This research was sponsored by the Carnegie Mellon School of Computer Science Keywords: PADO, Genetic Programming, Object Recognition, Evolution,Parallel Algorithms,Incremental Learning,Natural Images,Greyscale Video Images,Li...

Abstract not found

Genetic Programming is a method of program discovery/optimisation consisting of a special kind of genetic algorithm capable of operating on non-linear chromosomes (parse trees) representing programs and an interpreter which can run the programs being optimised. In this paper we describe a set of terminals and functions for the parse trees handled by genetic programming which enable it to develop effective image filters. These filters can either be used to highly enhance and detect features of interest or to build pixel-classification-based segmentation algorithms. Some experiments with medical images which show the efficacy of the approach are reported. 1 Introduction Genetic Programming (GP) is the extension of Genetic Algorithms (GAs) in which the structures that make up the population under optimisation are not fixed-length character strings that encode possible solutions to a problem, but programs that, when executed, are the candidate solutions to the problem [1, 2]. Pr...

The extraction of features for classification is often performed heuristically, despite the effect this step has on the performance of the classifier. The Evolutionary Pre-Processor is presented, an automatic nonparametric method for the extraction of non-linear features. Using genetic programming, the Evolutionary Pre-Processor evolves networks of different non-linear functions which pre-process the data to improve the discriminatory performance of a classifier. In experiments performed on 9 real-world data sets, the Evolutionary Pre-Processor was able to pre-process the data to reduce the test set misclassification rate. The dimensionality of the data was decreased and those measurements not required for classification were excised. The Evolutionary PreProcessor behaved intelligently by deciding whether to perform feature extraction or feature selection. 1 Introduction A common step in Pattern Classification is the extraction of features from the original data, motivated by the red...

Abstract: Rule-based systems used for Optical Character Recognition (OCR) are notoriously difficult to write, maintain, and upgrade. This paper describes a method for using Genetic Programming (GP) to evolve and upgrade rules for an OCR system. The language of the evolved programs was designed such that human hand-coded rules can be included into the initial population in order to upgrade for a new font. The system was successful at learning rules for large character sets consisting of multiple fonts and sizes, with very good generalization to test sets. In addition, the method was found to be successful at updating hand-coded rules written in C for new fonts. This research demonstrates the successful application of GP to a difficult, noisy, real-world problem. 1.

Interest operators play an important role in computer vision. Depending on the type of the environment some features may prove to be more advantageous than others. Thus detection of interesting features has to be made adaptive such that the best features according to some measure are extracted. We are trying to evolve such feature detectors using genetic programming. In this paper we describe our results where the desired operator, which is a Moravec interest operator, is directly specified. We show that the problem is a rather difficult one. Only an approximation to the Moravec operator could be evolved using several sets of elementary functions. 1 Motivation Interest operators play an important role in computer vision [8]. They highlight points which can be found easily using simple correlation methods. They can be used to calculate accurate distance information and for map building [23]. However no interest operator is suitable for all types of environments. A mobile robot which ma...

Optimally controlling the timings of traffic signals within a network of intersections is a difficult but important problem. Because the traffic signals need to coordinate their behavior to achieve the common goal of optimizing traffic flow through the network, this is a problem in collective intelligence. We apply a hybrid of a genetic algorithm and strongly typed genetic programming (STGP) to the problem of learning control laws which optimize aggregate performance. STGP learns the single basic decision tree to be executed by all the intersections when deciding whether to change the phase of the traffic signal. The genetic algorithm learns different constants to be used in these decision trees for different intersections, hence allowing specialization based on differences in geometry and traffic flow. Preliminary experimental work shows that our approach yields good performance on a variety of network configurations and that it can evolve control laws which induce cooperation, communication, and specialization among the traffic signals.

The maintenance of the high voltage electricity transmission network in England and Wales (the National Grid) is planned so as to minimise costs taking into account: -- location and size of demand for electricity, -- generator capacities and availabilities, -- electricity carrying capacity of the remainder of the network, i.e. that part not undergoing maintenance. This complex optimization and scheduling problem is currently performed manually (computerised viability checks can be performed after the schedule has been produced). This paper reports work aiming to automatically generate low cost schedules using genetic algorithms (GA). So far: -- A small demonstration problem has been identified, -- A fitness function has been devised, -- To date work has concentrated upon devising a representation based upon "greedy optimizers", which combine permutation GAs with scheduling heuristics, -- The best of these heuristics has been incorporated in the QGAME genetic algorit...