In this project we first formulate a theory of
strategy prediction in real-time strategy games
based the notion that a players strategical op-
tions are dependent on the in-game assets avail-
able. We then test a number of common training
algorithms to train multi-layer perceptrons to
predict the in-game assets of a player at time n
given information about information about the
same players in-game assets at time m, m < n.
A novel feature selection technique based on
real-time strategy game design principles is in-
troduced and tested, to address the number of
possible input features that can be extracted
from in-game assets.
A series of experiments to evaluate the
strategy prediction performance are performed
on MLPs trained with the Backpropagation,
RProp, Genetic Algorithm and two Genetic
Algorithm hybrids using Backpropagation and
RProp.