Automated Learning of Rules Using Genetic Operators

Automated Learning of Rules Using Genetic Operators C. E. Liedtke, Th. Schnier, A. Bloemer Institut für Theoretische Nachrichtentechnik und Informationsverarbeitung Universität Hannover, Appelstr. 9A, D 3000 Hannover 1, Germany E Mail: liedtke@tnt.uni hannover.de Abstract. The configuration system CONNY permits the automated configuration of image analysis processes which includes the selection of the appropriate sequence of operators and the adaptation of the free parameters. The system uses explicitly formulated knowledge contents from a human image analysis expert coded as rules of a rule based system. In the present contribution it has been investigated if and to which extent the rules can be learned automatically. The approach which has been chosen is based on the selection and valuation of individual rules and on the manipulation and generation of new rules by the use of genetic operators. The advantageous capabilities of a learning approach using genetic operators is demonstrated. Keywords: Automated learning, genetic operators, explicit knowledge representation, knowledge based systems, rule based systems, system configuration, image processing, expert systems, CONNY 1 Introduction Opto electronic sensors have a key position in present and future automation tasks. The wide acceptance of these systems is still hindered by the fact that the systems are quite complex and that the user is usually not at the same time an expert on image analysis problems. For these reasons a research project is conducted at the Universität Hannover, which has the goal to investigate to which extent the configuration process, i.e. the adaptation of image analysis systems to different tasks and different scenes, can be done automatically. In this connection the configuration system CONNY (CONfiguratioN system) [1, 2] has been developed. The concept is illustrated in Fig.1. The configuration system CONNY composes from an available set of image processing modules an image analysis path and adapts the free parameters like threshold values, window sizes, etc.. The task of image analysis is described by the user by providing symbolic information on one hand and test images on the other hand. The basic configuration cycle is illustrated in Fig.1. The image analysis system in its current state of configuration processes a test image and obtains a result which usually differs from the ideal result determined by the task definition. The configuration system contains a knowledge base with knowledge contents, which are independent from a specific image analysis task. The knowledge contents are coded to a large extent as rules. The rules describe the modifications of the image analysis process which are necessary in order to modify the processed result in a specific direction. They refer to specific properties of the processed result, like the interruption of lines and the amount of texture in Fig.1. For this purpose the system fully automatically extracts a valuation of the processing results employing natural spoken

Image Analysis System Test Image Reference Description Task Description Inference Engine Processed Result Interruption of Lines: strong Amount of Texture:... weak Result Description Interruption of Lines Close Lines Texture Occludes Contour... Reduce Threshold Knowledge Base Ideal Result Learning Component Data Flow Modification Implication Fig.1. Concept of the Configuration System CONNY terms. The inference engine combines the different knowledge contents, the requirements from the task description, the reference description, and the present description of the processed result and concludes how the image analysis system has to be modified in order to achieve a better processing result in the sense of the task definition. For the addition of knowledge contents about the configuration of new operators and knowledge contents about new task domains a learning component becomes necessary. Purpose of the present investigation was to find out to which extent knowledge contents represented by rules can be learned automatically. As area of application that part of the configuration system has been chosen, where rules are employed to modify parameter values based on the description of processed results in the parameter adaptation cycle. The learning method which has been selected is based on the use of genetic operators and will be presented in the following. 2 Learning Principle Traditional learning methods from the area of Artificial Intelligence are based on logical deductions. They have frequently been developed in order to simulate human reasoning processes. For special practical applications a major problem is caused by the fact, that

they require a vast amount of background knowledge similar to that of human beings. Since this background knowledge is usually not available an experimental approach has been used here, a genetic learning algorithm: New rules are generated directed and randomly and are tested. Rules which are successful form the basis for new rules, unsuccessful rules are eliminated. In this connection learning refers (a) to the assignment of credit to individual rules and (b) to the directed and random construction of new rules. New rules are constructed by combination of successful rules such, that success can be expected with a high probability as well. A main difference to traditional approaches is that new knowledge contents are not only made explicit from a knowledge base where they have been stored in an implicit fashion before, but that here real new knowledge contents can be generated on the basis of testing. The learning principle of the method used has been earlier described by Holland [3]. It is based on the concepts of the sensor actor model and the genetic pool and will be presented in the following. The sensor actor model (Fig.2) describes the embedding Environment Sensors Informations Rules Actions Actors Control Mechanism Fig. 2. Sensor actor modell of a rule base in an application. A rule consists of a condition part (IF...) and an action part (THEN...). The condition parts of the rules are compared with the information provided by the sensors. The actions in the action parts which are executed cause changes in the environment, which are then again sensed by the sensors. The learning and extension of the rule base is realized by the control mechanism indicated by the dotted line in Fig.2. A prerequisite for the learning is the existence of a global measure for the valuation of the quality of the actions. The quality measure is extracted from the information which the sensors provide. The control mechanism is by enlarge independent from the problem domain. The concept of the genetic pool requires a pool of individual rules. The action parts can be applied, when the condition parts match the conditions of the environment, but only the strongest rule is executed. For this purpose a credit has to be assigned to each rule, which expresses the amount of strength. Dependent on the success of the rule (measured by a global quality criterion) the credit value of this particular rule becomes increased (in case of success) or decreased (in case of failure). Thereafter rules are selected where the condition and action parts are combined to new rules (cross over operation). The selection is achieved by a random process which is influenced by the credit value. In this connection rules with a high credit value are preferred. With a certain probability rules are randomly modified (mutation). In order to keep the size of the pool constant weak rules are erased (principle of selection).

The cross over operator produces two children rules from two parent rules. In each parent rule one point is selected by random choice where the rules are split into two parts. The children rules are created by cross over assembly of the parts. The mutation operator modifies individual rules by modifying a randomly selected condition of the condition part or a randomly selected action of the action part. 3 Realized Learning Procedure Genetic learning procedures in the literature use mostly rules where the condition and the action part are represented by bit patterns. In the current project this had to be modified in order to arrive at rules as used within the configuration system CONNY. Some important new issues which had to be implemented were a concept for a flexible coding of the condition and action part of the rules employing natural spoken terminology, an implementation of the cross over and the mutation operator for the type of rules mentioned above and the development of strategies for the generation of new rules. In the following the most important aspects of the realization are described. A more detailed description has been provided by Schnier [4]. 3.1 Realization of the Sensor actor model In the application used here, the environment is represented by a set of test images and a processing path which produces contour images from the test images. The sensors are represented by routines which extract from the contour images a symbolic description. The description uses evaluation terms in a natural spoken terminology like Texturedness, Thickness, Intermingledness, Interwoveness, etc. The actors have the task to modify the parameters of the image processing path, like for instance lowering a threshold value. After having modified the parameter values the image analysis system computes a new contour image and modifies in that way the environment. 3.2 Coding of the Rules The coding employs the following principles: 1. Condition and action part are coded separately. 2. The comparisons in the condition part have the structure Operator Value Sensor. 3. The condition part consists of an arbitrary number of comparisons which are connected by logical AND operations. 4. The action part consists of one arbitrary operation and an actor which executes the operation. 3.3 Credit Assignment to Rules The valuation of a rule is expressed by the value of its assigned credit. The value with the label STRENGTH is related to the strength of the rule. It is recalculated any time the rule has been selected from the conflict set and the action operator has been executed. STRENGTH is calculated recursively using a factor ALPHA<1 for the new time instant t+1 from the previous time instant t to become

STRENGTH(t+1) = ALPHA * STRENGTH(t) + PREMIUM The values of STRENGTH(1) and ALPHA have been chosen on an experimental basis. PREMIUM is a monotoneous function of the adaptation gain which has been achieved during the adaptation step under consideration. The sign of PREMIUM is positive, if the processed result has improved and negative in case the processed result has become worse. 3.4 Implementation of the Cross over operator The cross over operator is applied to the condition parts of the parent rules only. The conditions of the parent rules are randomly distributed among the new condition parts of the children rules and the action parts are copied. The lengths of the condition parts of the new rules may differ from those of the parent rules. In order to prevent the combination of conditions, which are redundant or which exclude each other, the condition parts are modified such, that each sensor appears only once. The new rules are only added to the pool of rules, when they are not already existent. 3.5 Implementation of the Mutation operator The mutation operator can modify rules in two different ways. In contrast to the cross over operator the action part is included. First on a random basis a decision is made whether the condition or the action part is modified. If the mutation is applied to the condition part the location of the mutation is determined in addition by random choice. In a further random decision it is determined if the complete condition or action under consideration is replaced or if only an individual operator in the condition or action part is replaced. 3.6 Learning and Adaptation Sequences A task of the image analysis system during the phase of low level image processing may be the extraction of a contour image from a gray level image. As part of the processing a binarization operator has to be applied. In this application the system is as goal of the learning process supposed to learn how to adapt the threshold of the binarization operator. The learning of the rules is realized by using two interlocked cycles. The outer cycle has the purpose to provide the system with new tasks and to adapt the system to these tasks. One completed outer cycle represents one step of the learning sequence. Each inner cycle constitutes a completed adaptation. For the provision of a new task for each outer cycle a selection by random choice is made for (a) the test image from which contours have to be extracted, (b) the sequence of image processing operators in the processing path, (c) the values of the parameters of the operators including the initial value of the binarization threshold which has to be adapted. Using the chosen processing path the image is processed and the result becomes evaluated. If at this point an adaptation is not meaningful because the result is too close to the optimum or the result is too poor to be evaluated by the existing sensors, then the above mentioned steps are repeated. The inner cycle represents a simplified adaptation. From those rules where the conditions in the condition part are fulfilled, i.e. the rules in the conflict set, one rule is quasi

randomly selected for execution. The selection strategy uses a probability function proportional to the strength of the rule. The action is executed and the parameter value is modified. The image is processed again with the new parameter settings and the result is again evaluated. Dependent on the amount of success the strength of the rule is increased or decreased. With a certain probability (we experimented with 1:30) a genetic operator is executed. The adaptation cycle is repeated until no improvement becomes measurable or if 7 iterations at most have been completed. In addition the inner cycle is then terminated, when the set of instantiated rules is empty. Fig.3 shows an example of the sequence of adaptation steps within one inner cycle. Test image Start Quality 0.146 1. Step Quality 0.217 2. Step Quality 0.408 3. Step Quality 0.725 4. Step Quality 1.0 Fig. 3. Sequence of Adaptation Steps Using Learned Rules Fig. 4. shows the learning sequence depicted as the adaptation gain over the sequence of adaptations using Strength as credit only. An unsuspected drop of the adaptation gain to almost zero at a high number of adaptations becomes apparent. The reason is the following: The learning process develops rules which are on one hand very successful and on the other hand are executed only very rarely. They obtain a high strength as credit. Other rules which are used more frequently but are not as successfully obtain a low strength as credit. With increasing number of adaptations the number of rarely firing experts of high strength increases, while in order to keep the size of the pool constant the number of frequently firing non experts decreases. It decreases to that point that the conflict set becomes frequently empty, the adaptation cycle has to be terminated and the adaptation gain becomes then small or even zero. In order to overcome this problem the growth of the expert population had to be controlled. This has been achieved by including the criterion of firing frequency during credit assignment. It is used for the selection of those rules, which will become parent rules in connection with the genetic operations. For the selection of the rules in the conflict set the strength is still further used. With these improvements test with more than 16000 adaptations have been made. The result is shown in the upper curve in Fig.5.

0.00 Adaptation Gain 0.05 0.10 0.15 0.20 0.25 0 2500 5000 7500 10000 12500 Adaptations Fig. 4. Learning Based on Genetic Operators Using Strength as Credit The learning gain approaches fast a high level and increases then slowly but monotonously further. 4 Results For the evaluation of the implemented algorithm the learning performance of the system has to be assessed. For this purpose in each adaptation step the processed results after and before each completed adaptation are compared. The difference, the adaptation gain, is a measure for the improvement which has been obtained in this particular adaptation step. In Figs.4 and 5 the adaptation gain is depicted as a function of the number of adaptations during the overall learning process. The adaptation gain is strongly influenced by random effects like the values of the randomly chosen starting parameters. These parameters decide by enlarge which gain can be achieved at all. In some instances there is even a decrease in image quality after adaptation possible. Therefore, averaging over many adaptations becomes necessary to measure the true gain. Each plotted value in the representation of Figs.4 and 5 represents an average over 500 gain values. Further, a smooth curve has been added manually in order to enhance the graphical presentation of the general tendency of the adaptation gain. The numerical value representing the quality of the processed image has a range from 0.0 (very poor) to 1.0 (optimum). The randomly selected starting parameters produced in the average a quality of 0.46. Therefore, the adaptation could produce on the average a gain of 0.54 at most. The goal of the learning process was to learn rules how to adapt the threshold value which represents just one of several parameters in the processing path. All the other parameter values have been randomly chosen. The maximum gain during an adaptation can therefore only be obtained by modification of the threshold value alone, when the other parameter settings were by chance correct. Therefore, the maximum gain had to be expected to be on the average below 0.54. Fig.5 shows, that a gain in the order of 0.25 has been obtained which represents a very good learning success. Fig.5 shows the results of two learning algorithms, where in one the use of genetic operators is activated and in the other genetic operators are deactivated. In both cases the

Adaptation Gain 0.00 0.05 0.10 0.15 0.20 0.25 0.30 with genetic operators without genetic operators 0 4000 8000 12000 16000 Adaptations Fig. 5. Influence of Genetic Operators on the Performance of Learning learning has been applied to an initial pool of 300 randomly produced rules. During learning without genetic operators the gain has been achieved by development of a strength hierarchy only. A comparison of both methods indicates that by use of the genetic operators the gain could be doubled. References 1. C. E. Liedtke, A. Bloemer, Th. Gahm: Knowledge Based Configuration of Image Segmentation Processes, International Journal of Imaging Systems and Technology, Vol. 2, 285 295 (1990)69. 2. C. E. Liedtke, A. Bloemer: Architecture of the Knowledge Based Configuration System for Image Analysis CONNY,11th IAPR, International Conference on Pattern Recognition, den Haag, 1992. 3. John H. Holland: Escaping Brittleness: The possibilities of General Purpose Learning Algorithms Applied to Parallel Rule Based Systems, Machine Learning II, Morgan Kaufman Publishers Inc, Los Altos, 1986. 4. Th. Schnier: Untersuchung der Eignung genetischer Algorithmen zum Erlernen hochsprachlicher Regeln für die wissensbasierte Konfiguration von Bildanalyseprozessen, Diplomarbeit am Institut für Theoretische Nachrichtentechnik und Informationsverarbeitung der Universität Hannover, April 1992. Acknowledgements The project has been supported by a grant of the Deutsche Forschungsgemeinschaft.