Classifier system

Custer, L. L, Durham, S. K., Pearl, G. M. Probabilistic neural network multiple classifier system for predicting the genotoxicity of quinolone and quinoline derivatives. Chem. Res. Toxicol. 2005, 18, 428-440. 

Classifier systems are software tools that can learn to control or interpret complex environments without help from the user. This is the sort of task to which artificial neural networks are often applied, but both the internal structure of a classifier system and the way that it learns are very different from those of a neural network. The "environment" that the classifier system attempts to learn about might be a physical entity, such as a biochemical fer-mentor, or it might be something less palpable, such as a scientific database or a library of scientific papers. 

Classifier systems are in some respects a scientific solution looking for a problem because they are currently the least used of the methods discussed in this book. However, their potential as a disguised expert system is substantial and they are starting to make inroads into the fields of chemical and biochemical control and analysis. 

This software model is a learning classifier system. Because classifier systems learn, they can be applied to the control of a dynamic system, such as a reactor or an instrument, which must process various types of samples under unpredictable conditions, even when the rules required for successful control are unknown. 

As we shall see in this chapter, the classifier system is given no advance knowledge to help it plan what to do. Data on the chemical reactions that may be occurring or the values of rate constants are not provided therefore, it is wholly dependent for its success upon learning, which it manages by investigating the environment. This is a very different approach from one that relies on a comprehensive theoretical model of the reactions in the system. 

Despite these reservations, classifier systems have the potential to outperform model-based systems, particularly when theoretical models are weak. 

The most useful classifier systems learn without any direct help or instruction from the user. We shall meet these in the second part of this chapter, but first we consider how a CS programmed in advance could be used to control the temperature of a reacting mixture in a vessel. 

The classifier system, which analyzes the input messages, creates messages for output, and places them on an output list. 

The components that comprise, and the flow of messages within, a classifier system. 

In this discussion, we assume that the environment is a physical entity, but other environments may be used. It could be a database, a stream of messages, or any object that can provide input to the classifier system, accept output from it, and pass judgment on the quality or value of that output. 

This ternary coding has some limitations in particular, it may affect the ability of the system to derive general rules. The coding described in this chapter, which relates to work by John Holland, has been successfully used in a variety of applications and has the advantage of simplicity. Readers who wish to explore alternative ways of representing classifiers will find them described in recent papers on classifier systems. 

The CS comes into its own when the rules that are needed to control the environment are only partly known, or are completely unknown, so that a comprehensive set of ES rules cannot be constructed by hand. If we cannot create the ES by hand, it must also be impossible to create the CS by hand some other method of creating the classifiers must be found. This is the realm of the learning classifier system (LCS) in which all classifier systems of value lie. 

The steps in the combined genetic algorithm-classifier system are given below. 

A helpful starting point for further investigation is Learning Classifier Systems From Foundations to Applications.1 The literature in classifier systems is far thinner than that in genetic algorithms, artificial neural networks, and other methods discussed in this book. A productive way to uncover more... 

The rates of many enzyme reactions are strongly dependent on both pH and temperature. Construct a CS that learns to keep conditions within a simple reactor within the limits 6 < pH < 9 and 27 < T < 41. You will need both to write the classifier system itself and a small routine that represents the environment. Test the operation of your system by including a method that periodically adds a random amount of acid or base, or turns on a heater or chiller for a short period. 

Lanzi, P.L., Stolzmann, W., and Wilson, S.W., (Eds.) Learning classifier systems From foundations to applications, Lecture Notes in Artificial Intelligence 1813, Springer, Berlin, 2000. 

---