We have the numbers 0,1,2,3,4,5,6,7,8 and 9. But what are the most accurate or best shape of the numbers? That is. You don’t know exactly what a number looks like. But you want to use calculations to find out.

So my idea is would be to use machine learning and an evolutionary algorithm to generate the shape of the numbers.

The idea is to use unsupervised learning for a set of number images that start with random pixels. The strategy is that the images gets the right label from different relations. That is. You could have the problem to cluster the images into odd and even. So it would happen that picture1 gets label ”odd” and picture2 gets label ”even” and so on.

Based on the miss classifications you get for every thinkable problem. You then use an evolutionary algorithm to evolve a better picture0 to picture9.

One such problem could be to use an ?active handwriting filter to see if the pictures produce good output. Also rotating the numbers could reveal that the typical 6, 9 problem and be a cause for the recognition and genetic algorithm to change the picture representations.

//Per Lindholm