A Possible Route To Solving Chess

[towforce]

The outline plan:

* find a way to represent chess knowledge simply on a graph

* do machine learning against the graph (apparently, machine learning against large graphs is a good way to learn knowledge - link)

* from the resulting trained NN, produce a multi-dimensional polynomial (some NNs will allow you to do this)

* the resulting polynomial is likely to be large - but the only part of interest is where it fits the knowledge graph, so, for each dimension in the polynomial, work out the range that has an effect on the knowledge graph

* for these parts of the polynomial's range, make a new, and hopefully simplified polynomial by curve fitting. The resulting expression needs to be compatible with the knowledge graph

* maybe this simplified polynomial can give us insight into what constitutes a good relationship between chess pieces. The rules might be reasonably simple - we just hadn't realised what they really are (a lot of games look a lot easier with insight - nim is an obvious example).

Speaking from experience, a puzzle is a lot easier to solve when you have the solution (when you have the solution, it's often easy to work out how to construct the solution).

This method of solving chess might not work, but I do think chess is solvable for two reasons:

1. the increasing number of draws as elo level gets higher strongly implies the game is a draw, and we're getting close to the point where a machine will be unbeatable

2. a fruit fly displays many complex behaviours, but its brain has only 125,000 neurons and 20 million synapses. This proves that it's possible to encode a lot of highly complex knowledge/behaviour into a small computer