Thursday, 3 October 2013

An example of transdisciplinarity: expert systems and evolutionary cladistics 3. Expert systems and artificial intelligence (A.I.)

My "Quick Identification Chart" for Coriosolite
staters (coins) that I drew in 1989.  It is a paper
expert system. (Click on image to enlarge)
When I drew my "quick identification chart" in 1989 for the Coriosolite coins I was studying, I had no idea that I had built an expert system. Its true identity was later pointed out to me by Robert Van Arsdell when he read the manuscript for my book. I got the idea to do such a thing after using the identification tips for Egyptian scarabs that Sir William Flinders Petrie had revealed in: Scarabs and Cylinders with Names. After discovering that I had made an expert system, I labelled it as such and built an html version for my web site in 1995.

One of the advantages of "reinventing the wheel" is that you will possibly do it differently and that might even be better. After reclassifying Coriosolite (Celtic, from Brittany) coins, I found that instead of one series of coins, I now had three. Each series was the product of a different mint (one of them was not even Coriosolite, but was from the Unelli tribe in Normandy).

My thesis was that Coriosolite coin dies were created using an evolutionary process that introduced, modified and replaced design elements, and also corrected ongoing compositional problems that were caused by such changes. This avenue of investigation turned out to be far more fruitful than I had imagined. The original purpose was to create a chronology of these coin dies that was accurate right down to each variety, instead of trying to be accurate down to groups containing varieties. The latter system is the standard method in numismatics for establishing classes. Its flaw is that it is a system that is imposed on the data, and thus contains much subjectivity.

A better method is to allow the data to present its own structures and then classify that data through these. When I applied my method to the data, I differed from the standard method by including all design elements to build a chronology. Through a network of overlapping changes to discrete parts of the design, the chronologies revealed themselves (I had started looking for a single chronology).  I also saw that many design elements were not part of the evolutionary process -- the numbers of hairs on the head depicted on the obverse, while fewer in the most early coins, later were adjusted for compositional reasons such as a lack of space and so on; the earliest had fewer because the artist was still new to the composition and had not evolved a tidier design. I culled these non-evolutionary features, and built the chronologies based only on evolutionary changes and introduced novelties that had neither evolutionary nor compositional  qualities. Counting hairs on the head was a subjective action -- the artist might never have done so. He might have just drawn hairs within a pre-defined space until he used up that space. His considerations on the matter might have been formed on matters of "tightness" and "looseness".  The numbers of hairs plotted in the chronology revealed nothing consistent, so I removed them. What I was left with was, in essence, a genome of the evolutionary process of each series. Below, is an example (for my "Series Y"):

Series Y "genome"

Clear evolutionary jumps can be seen on features 1, 2, and 8 which are the eyes and nose on the obverse, and the pony on the reverse. These obvious features had already been selected to define the classes in the old classification system. The other features are more defining, but we will cover these in a later part. What makes this a discrete series is that the majority of features cannot be tracked before or after the sequence. In the old system, all of the subsequent three series had been considered to be only one. The small numbers at the top and bottom are the specimen numbers. and the letters H to M are my "groups". I introduced these as easy guides to the chronology. They are essentially the same as the previous "classes" -- albeit far more numerous, but I make it very clear that these are subjective groupings and cannot thus be used in any statistical analysis of the primary data.

The important discovery was that there were no intentional groupings of the dies within their revealed chronology. This fact was also backed up through an examination of the specimen's weights and samples of the elemental content of the metal which had previously been revealed through neutron activation analysis in Katherine Gruel's: Le tresor de Trebry, Cotes-du-Nord, 1er siecle avant notre ere: Contribution a l'histoire du monnayage des Coriosolites : methodes physiques et ... l'Universite de Besancon) (French Edition). In other words, the embedded series Y had its next embedded arrangement in the specimens themselves -- there were no objective "classes".

In the Wikipedia entry for expert system, I find that I have issues with the very first sentence: "In artificial intelligence, an expert system is a computer system that emulates the decision-making ability of a human expert." For a start, I found a paper in an AI journal saying that expert systems had been removed from the subject of artificial intelligence because they worked, while AI did not. The same paper praised the absolute yes/no method (which I had used) and had criticized the sort that used fuzzy logic because of the subjective nature of its "weighting" -- for example, "cold" and "very cold" definitions vary person to person and are also dependent on any current situation. Of course, expert systems are not restricted to computers -- there are many paper systems being used -- a "choose your own adventure" being a well-known example. What raised my hackles was, "that emulates the decision-making ability of a human expert.". I suppose I could have tried to build such a thing -- virtually all subjective and modified by various parts of my life across a myriad of subjects -- it never occurred to me at the time to attempt such a silly thing. Nor did I base my expert system on the structure of the chronology --as on the Series Y chart although I, later, saw a few expert systems that had been so designed.

My method was to keep dividing the data into halves (or once thirds) and to find the fastest route from the first question to the answer that would have delivered by the expert using his own method. That is the method shown in the first illustration. I first split up the data between broad stylistic differences which largely marked the three different series in the coinage. This was done because the large number of variations could only be understood in each local part of the "genome". The structure I drew is also a cladistic "tree" and was the first application of cladistics to archaeology, being fourteen years before Cladistics & Archaeology recommended applying cladistics to archaeology.

 The process of designing the expert system also allowed me to see a potential problem in evolutionary cladistics, but that will have to wait until later. In the next episode I will make the necessary introduction to this transdisciplinary comparison.

No comments:

Post a Comment