Task-Driven Species in Evolutionary Robotic Teams - Foundations on Natural and Artificial Computation - page 143

- The n coecients of the control system. Each robot has a control system rep-
resented by a RBF ANN with 9 inputs: the distance and angle to the nearest
inactivated block, activated block and assembled block, and the robot speeds
s assembled , s isolated and s empty . The ANN has only one output providing the
motion angle of the robot. The number of coe cients n depends on the
selected number of ANN hidden layers and neurons per layer.
Once the Individuals and Environment components have been defined, imple-
menting this experiment in ASiCo implies defining the chromosome represen-
tation, the energy flow strategy and the interaction set. As commented above,
the chromosome is made up of the n ANN coecients, the three coecients
that establish the effective robot speed, two grasping parameters and the three
vision range parameters. Consequently, the ASiCo evolution has to control both
morphological and behavioral parameters of the robots. The energy flow strat-
egy has been defined as follows: each robot is born with 10000 energy units and
starts consuming one unit on each time step when it reaches the age of 1000
(representing a mature age). The task is considered to be accomplished every
time an assembled block is placed in a collection area. This action provides 2000
energy units to the robot or robots that have participated in the task, that is,
to each robot that has found a block, joined two blocks or carried the assembled
block to the collection area. When a robot runs out of energy, it dies. As can be
seen, the robots increase their private utility when accomplishing the task indi-
vidually, being the global utility maximization, the one that guides the system
towards the completion of the task, a consequence of the maximization of the
private one. Finally, creating the ASiCo interaction set implies establishing the
rules that define the actions, reactions and perceptions of the robots and the
blocks. They are all very simple, and constitute the scenario rules.
The final objective of the multirobot system is simple: collecting the maximum
number of assembled blocks (dark rectangles). Hence, previous to the collection
of these blocks, the robots must create them by assembling two activated ones
(light squares), which must be found by the robots in their inactivated state
(dark squares), as displayed in Fig. 2. The task has been designed to allow spe-
cialization through the adjustment of the chromosome parameters to exploring
the environment, carrying activated blocks or carrying assembled blocks. In this
sense, an optimum way to accomplish the task would be using three different
species. Obviously, this is not the only option, being the different specialization
possibilities the main characteristic of this experiment. Anyway, an homogeneous
solution would be suboptimal in this case because the robots that adjust their
parameters to intermediate values for the three commented subtasks, will not
perform any of them properly.
Experimental results: with the previously presented scenario and elements,
ASiCo evolution has been executed several times with a population size of 20
robots during 150000 time steps. The ANN applied had one hidden layer with
9 neurons, implying that 120 coecients were included in the chromosome to
define it. In this section, we are going to present the typical results obtained for