Glasgow Genetic Algorithm Demonstrator Please wait for a few moments while EA_demoTM below is loaded for interaction. For the EA GA Demo, you'll need JavaTM runtime (download here: java.com). NB. Java applets may not run in IE8 that turns DEP on by default. EA and CAutoD interactive demo - run Java applet below Learn step by step, or convince convergence with background batch.  See how x and y jump out of local optima.  Experiment with initial population, mutation rate, crossover rate, selection mechanism, elitism, and constraints.  Gain more sense ...

Further Reading - Some SCI Indexed Papers on Google Scholar:

• Book: Real-World Applications of Evolutionary Computing, S Cagnoni, R Poli, & Y Li, et al, Springer-Verlag Lecture Notes in Computer Science, 2000, 396 pp, Volume 1803/2000, Berlin, ISSN 0302-9743, ISBN 978-3-540-67353-8, DOI 10.1007/3-540-45561-2.
• Book: Parallel Processing in a Control Systems Environment, E Rogers & Y Li, Prentice Hall Series on Systems and Control Engineering, 1993, 364 pp, ISBN 0-13-651530-4.
• Some recent papers can be downloaded here from the University's eLibrary - Enlighten

From CAD to CAutoD for Engineering Systems

To meet the ever growing demand in quality and competitiveness, a ‘good design’ of an engineering system needs to meet multiple objectives such as energy efficiency, maximal output, highest speed and cost-effectiveness. The engineering design problem concerns both finding the best design within a known range (i.e., through 'learning' or 'optimisation') and finding a new and better design beyond the existing ones (i.e., through creation and invention). This is equivalent to a search problem in an, almost certainly, multidimensional (multivariate), multi-modal space with a combinational design objective or multiple objectives.

If the objective function (or, inversely, cost function) is differentiable under practical engineering constraints in the multidimensional space, the design problem may be solved easily by setting its (first) vector derivative to zero. Finding the parameter sets that result in a zero first-order derivative and that satisfy the second-order derivative conditions would reveal all local optima. Then comparing the values of the performance index of all the local optima, together with those of all boundary parameter sets, would lead to the global supremum, whose corresponding "parameter" set will thus represent the best design.

Unfortunately, the above scenario does not exist in engineering practice. At present, many designs and refinements are mainly made through a manual trial-and-error process with the help of a CAD simulation package. Usually, the adjustments need to be repeated many times until a ‘satisfactory’ or ‘optimal’ design emerges.

This adjustment process could be automated by computerised ‘intelligent’ search. For this, an evolutionary algorithm (EA) based multi-objective "search team" may be interfaced with an existing CAD package in a batch mode. An EA encodes the design parameters in binary or integer strings and varies the strings to refine multiple candidates through parallel and interactive search. In the search process, it selects better performing candidates using ‘survival-of-the-fittest’ learning. To obtain the next ‘generation’ of possible solutions, some parameter values are exchanged between two candidates (by an operation called ‘crossover’) and new values introduced (by an operation called ‘mutation’). This way, the evolutionary technique makes use of past trial information in a similarly ‘intelligent’ manner to the human designer. The EA based optimal designs can start from the designer’s existing design database or from an initial ‘generation’ of candidate designs obtained randomly. A number of finally ‘evolved’ top-performing candidates will represent several automatically designed optimal motors.

For experiencing the working mechanism of a coding EA (i.e., a genetic algorithm) and how it enables design automation, put your hands on the above interactive and animated GA demonstrator. This is being further improved and enhanced by two project students.

Other Elements for NEC4

Conventional computers

Genetic evolution - A way to solve complex optimization problems

• The complexity of a problem lies in the complexity of the solution space that may be searched through. This complexity arises due to:

1. size of the problem domain;
2. non-linear interactions between various elements (epistatis);
3. domain constraints;
4. performance measure with dynamics and many independent and codependent elements; and
5. incomplete, uncertain, and/or imprecise information.

• Systems of nature routinely encounter and solve such problems. Good examples include genetic evolution of species, which become better and better suited to the environment generation by generation.
• A Darwinian machine or an evolutionary program emulates this process and does not require an explicit description of how the problem is to be solved. It tends to evolve optimal solutions automatically.
• Evolutionary computing technology:

1. Maps an engineering system directly onto a "genetically" encoded (numerical or character) string of system parameters and structures;
2. Can convert an automatic design problem to a simulation problem to solve;
3. Solves a difficult design problem by intelligently evaluating performance and evolving optimal candidates based upon the evaluations.
4. A universal tool for design automation;
5. Also applicable to the design of ANNs, called Darwinian selective learning;
6. Such a nearly-untapped powerful technology that will make a revolutionary impact on engineering design in the near future.

Learning and parallel solutions through artificial neural networks

• Learning to drive a car = Training your neural network
• Conventional artificial intelligence lacks learning and adaptation features
• The need of learning and parallel processing
• The solution is neural networks:

1. A neural computer adapts itself during a training period, based on examples of similar problems often with a desired solution to each problem (just similar to automatic control).
2. While in application, it can also adapt itself to a brand new problem and is able to offer a viable solution.
3. An ANN behaves as a universal functional approximator.
4. Major drawback of an NN - mapping between the problem (or engineering system parameters) and the network parameters is indirect and, thus, so is learning.

• Evolving ANNs for target applications:

1. Design difficulties have held back the applications, as seen in history.
2. Are design techniques mature today? No, different types of NNs (e.g. feedforward vs. feedback) require different learning algorithms and there are different types of learning (supervised, unsupervised and hybrid).
3. Existing learning algorithms are difficult to apply to some real-world problems.
4. Some engineering problems still difficult to solve, such as the direct application of ANN for embedded control (i.e., use as a normal feedforward path controller).
5. Let alone training of "irregular" networks and developing efficient architectures.

• Application examples: Daimler-Benz, British Gas, DTI’s campaign

Fuzzy logic and its embedment in a neural network

Evolutionary Computation at Glasgow: Yun Li's Intelligent Systems Group.  The materials on this page were initially provided for use with University of Glasgow's Neural and Evolutionary Computing IV (1995) course. Page written by Yun Li (李耘), 1997 (So please don't ask me for source code - it's long buried under the pile).