Evolutionary Neural Networks: Design Methodologies
By ai-depot | March 24, 2003
Conclusion
It has to be noted that the same input-output mapping can be implemented by different neural network architectures. Given a problem in hand, the topology for an ANN is not unique. Genotype representation of two structurally different ANNs will be different even though the functional mapping they define may be same. EAs are not able to detect these symmetries and hence a crossover in such a case would very often result in an unviable offspring. The search space is drastically increased and the efficiency of the operators is also severely affected. Moreover, in networks where more than one task needs to be learnt, there are chances of incompatible roles getting combined leading to similar problems. A simple solution to these problems would be to restrict the selection operator to smaller populations, and to introduce intuitive biased measures in crossover and mutation.
ANNs are capable of learning very high-order statistical correlations that are present in a training environment. Learning algorithms provide a powerful mechanism for generalizing behavior to new environments. However, the purpose of the networks then merely reduces to the simulation of a set of goals. Networks in which endogenous goals plays an important role in determining behavior are difficult to train using usual training algorithms. In such cases evolutionary methodologies are the appropriate mechanism for developing goals and purposeful behavior, rather than goal-directed explicit training.
The notion of an “optimal solution” in EAs is a bit deceptive. Fitness evaluations are all done with respect to the current population only; hence a good solution is better only in comparison to the members of the current population. There exists no way to check how good a solution is in the overall search space. The end conditions are also not explicit, and so the quality of the solution has to be a compromising act.
Global search methods like EAs are usually computationally expensive. They are more advisable to be used when little or no prior information about the network is available, and the performance value required out of the ANN is high. More effectively, they are good algorithms to start with the design and once some knowledge is gained, other purposive algorithms can come up with the solution faster. Parallel implementations of these algorithms will also become more and more purposeful as the need for designing real world applications arises. As the idea of natural adaptation become more and more promising, EAs and ANNs can serve as quite effective tools in narrowing the gap between simulated and adaptive behavior. The direction of learning is going to be the one from which it can obtain the best advantage.
Written by Rinku Dewri.
Tags: evolutionary algorithm, genetic algorithm, neat, neural network
Category: tutorial |
