next up previous contents
Next: A Framework for GBML Up: Introduction Previous: Machine Learning   Contents

1.2 Arguments For and Against GBML

GBML methods are a niche approach to machine learning and much less well-known than the main non-evolutionary methods, but there are many good reasons to consider them.

Importantly, the classification accuracy of the best evolutionary and non-evolutionary methods are comparable [95] §12.1.1.

Synergy of Learning and Evolution
GBML methods exploit the synergy of learning and evolution, combining global and local search and benefitting from the Baldwin effect's smoothing of the fitness landscape §2.3.

There is some evidence the accuracy of GBML methods may not suffer from epistasis as much as typical non-evolutionary greedy search [95] §12.1.1.

Integrated Feature Selection and Learning
GBML methods can combine feature selection and learning in one process. For instance feature selection is intrinsic in LCS methods §3.5.

Adapting Bias
GBML methods are well-suited to adapting inductive bias. We can adapt representational bias by e.g. selecting rule condition shapes §3.5.2, and algorithmic bias by e.g. evolving learning rules §3.4.

Exploiting Diversity
We can exploit the diversity of a population of solutions to combine and improve predictions (the ensemble approach §3.3) and to generate Pareto sets for multiobjective problems.

Dynamic Adaptation
All the above can be done dynamically, to improve accuracy, to deal with non-stationarity, and to minimise population size. This last is of interest in order to reduce overfitting, improve run-time and improve human-readability.

Evolution can be used as a wrapper for any learner.

Population-based search is easily parallelised.

Suitable Problem Characteristics
From an optimisation perspective, learning problems are typically large, non-differentiable, noisy, epistatic, deceptive, and multimodal [207]. To this list we could add high-dimensional and highly constrained. EAs are a good choice for such problems.

See [62] and §3.4 for more arguments in favour of GBML. At the same time there are arguments against using GBML.

Algorithmic Complexity
GBML algorithms are typically more complex than their non-evolutionary alternatives. This makes them harder to implement, harder to analyse, and means there is less theory to guide parameterisation and development of new algorithms.

Increased Run-time
GBML methods are generally much slower than the non-evolutionary alternatives.

Suitability for a Given Problem
No single learning method is a good choice for all problems. For one thing the bias of a given GBML method may be inappropriate for a given problem. Problems to which GBML methods are particularly prone include prohibitive run-time (or set-up time) and that simpler and/or faster methods may suffice. Furthermore, even where GBML methods perform better the improvements may be marginal.

See the SWOT (Strengths, Weaknesses, Opportunities, Threats) analysis of GBML in [224] for more.

next up previous contents
Next: A Framework for GBML Up: Introduction Previous: Machine Learning   Contents
T Kovacs 2011-03-12