4play is a new program that performs neuroevolution. It uses
simulated genetic evolution in order to optimize an ANN. 4play is
being developed by the opensource project ANNEvolve. It is written in
standard C for all platforms, and may be downloaded from the
sourceforge website, below.
The motivation for this software is to demonstrate that an ANN can
store data in memory, and make use of that data. This is something
that cannot be done by the most common types of ANNs, such as MLP's.
Feed-forward ANNs have only the equivalent of ROM; they have no
writeable memory. Training or evolving the ANN determines a set of
weights and biases that can handle a given problem. This memory is
analogous to ROM. It is also analogous to species learning, as done
by ants or bees, whereby the species as a whole has learned much, but
the individuals learn little. In contrast, there are problems that
require learning after the training or evolution is completed. In
principle, an ANN with recurrent connections can act as a RAM and
store information even with fixed weights & biases. The challenge was
to create an ANN that "knows" how to make use of this ability.
Therefore a problem was chosen in which active, RAM-like, memory was
central to the solution. The 4-card puzzle in 4play requires the ANN
to remember the locations of two cards, and also the identity of those
cards. Then it must make correct decisions based on that information.
This puzzle is loosely based on the game of Concentration, where cards
are spread face down on a table and the players get to turn over two
cards, trying to form pairs. Each turn starts with the cards face
down, since the main point is to remember what cards were exposed, and
where they are located.
This is the problem to be solved by an ANN: There is a deck of 4
cards, 2 aces & 2 deuces. The cards are shuffled and dealt face down
on the table. This pattern of 4 face down cards is presented to the
ANN. Then one card is exposed and the new pattern is presented to the
ANN. A second card, not matching the first, is then exposed, and this
pattern is presented. Then the cards are turned face down again, this
pattern is presented to the ANN, and this time the ANN makes a choice
of a card to be exposed. The chosen card is exposed and the 4-card
pattern presented to the ANN. It then makes its 2nd choice to
conclude the game. If its two choices form a pair, it wins.
Otherwise it loses.
The goal is to evolve ANNs that are perfect players of this puzzle.
This has been achieved. The 4play package includes "chromosomes" that
create perfect ANNs when read into the evolving population.
Much more information about 4play is included in the downloadable
4play is extensively
The Annevolve team is experimenting with the Evolution of Artificial
Neural Networks (ANNs). They are combining the two fields of
Artificial Neural Networks and Evolutionary Computation. Find out
more about the other neuroevolution programs at:
- posted 18 years ago