Game Development Reference
In-Depth Information
Of course, ''better'' can have many different interpretations: easier to program,
produces a better result, can be done faster, etc.
The trick with teaching—or training , as it often is called—is to make sure that
the right things are taught. This is not always a given. A system trained to
distinguish men from women at first had trouble with long-haired male rock
stars. Another image-analysis system learned to distinguish the photographic
qualities of the training data instead of differences in the subjects of the pho-
tographs. The system is only as good as the training data; it will faithfully reflect
any biases in the training data. Lionhead Studios improved the user interface to
the learning system in Black & White II to make it explicitly clear to players
exactly what feedback they were about to give their creature, solving a major
source of frustration with the original game.
The most straightforward way to ensure that learning has taken place is to
administer a test with problems that the student has never seen before. This is
called an independent test set . A system that is tested using the same data that
trained it is subject to memorizing the problems instead of working out the
solutions. Just as other software is tested, learning systems need test data that
covers a broad range of the typical cases, as well as the borderlines and outliers.
The system is proven to work only in the areas that the test set covers. If the
system will never encounter a situation outside of the training set, the demand
for an independent test set is relaxed.
Hidden in this coverage of training is a fact that needs to be made explicit: We are
talking about two different kinds of training. The two methods of training are
supervised training and reinforcement training. In supervised training, the
learning algorithm is presented with an input situation and the desired outputs.
The image-analysis system designed to identify men versus women mentioned
earlier was taught this way. The system trained against a set of pictures of people,
where each picture was already marked as male or female. Supervised training
happens before the system is used and not when the system is in use. If there is a
knowledgeable supervisor available at run-time, there is no need for the learning
system to begin with. The success of supervised training in learning algorithms
outside of the game industry has not been echoed within the game industry.
Unless otherwise noted, supervised learning is the default when discussing
learning algorithms.
Search Nedrilad ::

Custom Search