CARACTERÍSTICAS GENERALES

The paradigm of learning and performing highlights 2 important points. Firstly, “excess energy” and “excess time” must be present to facilitate non-critical periods where learning can occur. If the best known performance is always demanded of a system, then there is no opportunity for variation from the current best solution and hence learning cannot occur. The second important point is that performance is not the same as fitness. Perfor- mance is the perceived fitness over a period where the system behaviour matters to the perceiver.

For all problems we can make the very general claims that:

• a problem must come into existence at some point in time or is identified at some point in time

• once the problem is known it will take some finite amount of time to devise a solution

• once an acceptable solution is believed to have been found, it could be applied immediately or further time could be spent devising an even better solution

There will be a class of problems where it makes most sense to generate solutions by interacting with the world directly (i.e. in situ) and not via a simulation of the world (i.e. in silico). Therefore, regardless of whether a solution is devised offline or online, the real-world will be used for feedback during the generation of the solution. Providing the same solution generating mechanism is used, then the time to generate an acceptable solution should be the same for either approach. We will continue to employ evolutionary algorithms (though not evolutionary programming) as the solution generating mechanism. For the problems used in this analysis we can think of in situ evolution being equivalent to

requiring each evaluation taking a finite period of time and, most importantly, evaluating a solution affects the success rate of the system when FI is greater than zero.

Let us suppose that a problem presents itself at some specific time, g0. For conve-

nience we will quantise time into discrete periods, namely generations, corresponding to the period of evaluation of the population for the respective generation. Additionally, let gi

correspond to the point in time immediately after the completion of the evaluation of the ith_{generation. Hence the period over which the first generation is evaluated is g = 1 which}

occurs between g0 and g1.

Let us additionally assume that we are keen to deploy a solution when it becomes available. More specifically, that we have the flexibility to wait for an acceptable solution to be evolved but not so much as to be able to wait for evolution to an optimal solution. Let gstart be this point in time after an acceptable solution has become available. Similarly,

let gf inish be the point in time when solving the problem is no longer of interest or the

importance of performance is once again zero (i.e. Φ(g) = 0 where g ⩾ gf inish).

Thus, for the given problem, we have a fixed period of utility from tstart and tf inish

defined by when Φ > 0. The importance of fitness during the period of utility will be lifted such that success over the time epoch of one generation requires the pool fitness to be between Fpool(gstart)and F_elite(g_start). This means that success is known to be achievable as it will less than the current known elite, but also not so trivial that any population can achieve success. The lower limit of Fpool(gstart)is a logical bound since in theory this will be the observed performance of the optimal learning population and so without any loss of learning potential, this level of performance can be achieved. Setting an acceptable performance target lower than this will tend to be trivial to achieve providing the fitness landscape remains constant.

For experiments where the FI function changes during the period of utility, the change will be limited to only one occurrence. Where the fitness landscape also changes, we will coincide the change of fitness importance to the same generation that the change in the fitness landscape occurred.

g

start

g

change

g

finish

F

it

n

e

ss

generation

elite ɸ = 0 pool ɸ = 0 pool ɸ = .5 elite ɸ = .5 pool ɸ = 1 elite ɸ = 1

F

accept

F

accept₂ 1

Figure 4.8: Hypothesis of how experiments will be evaluated. Showg0,gstart,gf inish, change

of landscape,Faccept1,Faccept2

benefit (in terms of success rate) over the offline and pure learning approaches. Practi- cally, we need to set a start time, even if the timing is somewhat arbitrary. In an attempt to construct a fair comparison, we will employ only one method for solution finding, namely genetic evolution, and allow all approaches to evolve for the same number of generations before the importance of fitness is increased from 0. This will occur at generation, gstart,

at which the importance will be lifted such that success requires a pool fitness that lies be- tween the Fpool(gstart)and Felite(gstart). Furthermore, gstartwill be set prior to convergence

In document Impacto social de los programas de pregrado de la universidad de la sabana (página 74-79)