Artificial life anthill simulation with visualisation.
Ants should collect food to maintain enough energy to survive,
they can reproduce, fight and move on an evolutionary island, and between them. During the
work of the system, current statistics of average coefficients of the ants such as energy are
displayed. Position of ants on the island is also visualised. Below a screenshot of the system
can be seen.
Evolutionary multi-agent system devoted to the problem of neural network
optimialisation. Evolutionary neural networks are used to solve the problem
of time series prediction.
Neural networks used in the experiments were multi-layer perceptrons
with three layers. There were 6 neurons in the input layer
and 15 neurons in the first and the second hidden layer.
The sigmoid activation function was used in hidden layers and
linear function in the output layer. The learning rates for hidden and
output layers, momentum, penalty function parameters and weight
cut-off threshold were the subject of evolution process.
The networks were trained with standard backpropagation method with momentum.
The results described below were obtained for Mackey-Glass time series
with delay parameter equal to 30 and step parameter equal to 10.
The signal range was from 0,2 to 1,4.
1. Original and predicted time series
2. Prediction error
In the graph (see fig. 2.) an absolute prediction
error of selected agents and global prediction error are presented.
The process tends to be faster at the beginning and slows down after
few hundreds of steps. PREMONN classification mechanism allows to keep
the global prediction error quite small after few dozens of learning
epochs, although prediction of particular agents may not be accurate.
3. Number of agents in the system
2. Average energy of an agent
A crucial task in these kind of systems is to maintain stable number
of the agents in the population in order to continue the evolution process.
In the conducted experiments the
population of the agents seems to be stable, as it can be observed in the
graph (see fig. 3.), where the number of agents in the
system is presented. It is to notify that (similar to average prediction
error) the number of agents at the beginning of the evolution process changes
very fast, then begins to stabilise.
As the processes of evolution and death are based on the life energy of
agents, it can be seen in the graph (see fig. 4.), that
amount of this resource is also stable during the operation of the system,
which proves that mechanisms of resource distribution agents (prices and
penalties) are satisfactory.
Evolutionary multi-agent system originating from EvNet but with
immunological mechanisms of agent evaluation introduced. The problem solved is
also time series prediction.
Energy management is based on existence of so-called B-Cells that are distributed
among the agents in the system. Every B-Cell evaluates neighboring agents
and changes its energy level.
System is currently under development, results will be shown shortly after
the first release.