Developing Games That Learn - Tapa dura

If you've ever wondered how to empower your software with the ability to learn from experience, this book is for you. Developing Games That Learn shows in detail how to implement single-trial learning, illustrating its approach in board games Tic Tac Toe and a more complex game called Drop Four. The authors discuss the learning algorithms and present the source code that implements them.

The program learning techniques are presented in sufficient detail to be useful for practical software design. They are particularly suited to computer game development, but can also be used in other areas.

This book provides valuable guidance for programmers, project leaders, and game designers.

Existing approaches to program learning are often of the "dumb" type -- the program simply learns not to repeat any sequence that led to failure. Authors Dorfman and Ghosh take a more powerful path: for each example of failure, their method first finds the point of no return -- the point after which the game was certain to be lost. Then that point is prevented from being reached via any of a number of different (but equivalent) paths, not just the single path actually played, thus accelerating the learning.

If you are interested in using these techniques in applications other than games, the authors discuss to do that.