Hi all, just wondering what type of problems GA''s are best suited towards solving and if a GA can ever be used to ''learn'' to play simple puzzle games like tetris? Also does anyone have any ideas of alternative puzzle games a GA can be used to play and learn? Thanks DarkStar UK ------------------------------- Loves cross-posting because it works

A genetic algorithm is basically a way of searching through a large set of potential answers to a problem. If you can represent numerous possible ways to play a game, and assign some sort of relative scoring to allow you to meaningfully compare one way to another, then a GA could be used.

[ MSVC Fixes | STL Docs | SDL | Game AI | Sockets | C++ Faq Lite | Boost
Asking Questions | Organising code files | My stuff | Tiny XML | STLPort]

The answer you will often find in the literature is that GAs are particularly suitable for solving blind search problems . These are problems in which an objective function (often known as the fitness function ) is not known analytically, or its derivatives are not known analytically (or are not computable). Hence, the only information available is the value of the objective function at discrete points in the state space. This is true of many problems where one can evaluate the ''fitness'' of a solution candidate without knowing how the objective function varies around that solution or any other solution. For example, you might, as fup often does, use a genetic algorithm to propose solution candidates that comprise the set of parameters for an artificial neural network. The ''fitness'' of the network might be computable as some function of the outputs given the inputs. Clearly we don''t know what the objective function is for the parameters, but we can evaluate the quality of the solution.

So, can you solve Tetris with a GA? Theoretically yes. Practically? That depends on how you choose to represent the problem.

I hope this in some small way helps your understanding of GAs and their applicability.

Cheers,

Timkin