Ultra fast cnn based Hardware Computing Platform Concepts for adas visual Sensors and Evolutionary Mobile Robots

 
103 
define any detail for the locomotion behavior. After generating new chromosomes, we 
obtain the corresponding fitness value by applying the fitness function. The main point on 
applying the fitness function is that this function is not a real time procedure and that the 
result from the fitness calculation will be only ready after a certain period of time beyond 
the time the wave effect will act on robot hinges/actuators. In fact, one cycle of time (i.e. 
one period of the wave acting on the actuator) is not sufficient for measuring with good 
accuracy the position in space of the robot. Many cycles of the wave generated are 
necessary to be applied to robot actuators. By measuring some robot parameters like the 
position of the robot central point, the robot angle (related to the global coordinates of the 
system) and so on, the fitness function is quantified. In the initial state of the training phase 
the algorithm selects some randomly generated chromosomes. There is no rule for 
evaluating how many chromosomes should be generated in the initial population. This 
number varies depending upon the complexity of the problem [149]. Some authors have 
defined 100 generations of chromosomes/genes for the initial state [149, 150]. After each 
generation, a fitness function is used to evaluate the cost of chromosomes in the simulator 
leading to maximum efficiency. During our computations, each evaluation took approx. 3 
seconds and the program spent approx. 60 seconds to evaluate appropriated 
chromosomes. After this step, both chromosomes and fitness values will be sorted with the 
aim/goal of minimizing the fitness values in a link list. The next step concerns the 
crossover (i.e. both selection and breed) of chromosomes. Our experiments have shown 
that 50% of the best chromosomes are fitting for the crossover. It was found that this range 
has a good probability to generating the better chromosomes. In each step, we randomly 
select 2 chromosomes in this range for the crossover process. Many evaluations have 
shown that the use of the “two
-
point” technique for the crossover is the best solution
. In 
this process we define two points (randomly) on the selected chromosomes; the contents 
of the chromosomes between these two points are exchanged (Figure 9-5). In Figure 9-5 
P1 and P2 are two randomly selected points. S1 and S2 are two selected 
parents
/chromosomes. The crossover leads to two new “children” (see Ch1 and Ch2 in
Figure 9-5) with new properties. During the trial and error process, we obtained that the 
good probability for mutation is around 10%. This rate is essential for avoiding the local 
minimum trap. In the long term, this rate of the mutation increases the quality of 
chromosomes in the list [151]. 

 
104 

Download 3,22 Mb.