RCRT GAVONTS: Genetic Algorithm Vector Optimised Nonlinear Transformations - is our genetic algorithm based 'meta-model' technology. This is one of the quantitative modelling and analysis services that we provide, it is aimed at tackling non-linear problems, Genetic Algorithm Nonlinear Transformations - DNA Image and situations where there may not be a suitable prototypical mathematical 'model' available.

The GAVONTS method is designed to find a nonlinear transformation that maps an 'input vector' of values onto an 'output vector', and which optimises a chosen set of criteria in the process.

For example, the 'input vector' might represent the state of a complex system; and the output vector could be an estimator for the future state of that system. The optimisation process would then correspond to improving the quality of a forecast of the system's behaviour.

GAVONTS uses a combination of Genetic Algorithms, together with a particular way of building and combining complex non-linear transformations that can be represented using a set of specially designed xml data structures.

Our Genetic Algorithm modelling framework then simulates the process of evolution, and finds a mathematical model (or models) that best solve the given problem. The 'genetic' element relates to the way in which new models are created from parent models. This approach does require that we have an appropriate way of 'breeding' mathematical functions; this aspect is something of a challenge.

RCRT GAVONTS technology has been selected for listing by the US Department of Defense's Data & Analysis Center for Software (DACS) - the DoD Software Information Clearing house, managed by the United States Air Force, serving as an authoritative source for state-of-the-art software. More...

GAVONTS - Applications

There are a wide range of potential applications for GAVONTS. In the financial arena alone, these range from credit scoring and insurance risk estimation, through to financial time-series forecasting and risk analysis. Other, more diverse applications include problems such as the detection of weaknesses in cryptographic systems.

Also see:

The GAVONTS web site. - GAVONTS now has its own dedicated web site.
'Lander Example' - in which GAVONTS has been used to create a control system algorithm that can successfully land a spacecraft in simulation - this includes a short virtual reality movie clip showing such a descent.
The XML-Space GAVONTS Example page - which also discusses this technology.

For further information, contact:

Modelling Complex Problems

When we 'model' systems, and fit data to these models, there are two basic approaches:

We understand the internal dynamics of a system well enough to create a bespoke model that represents its behaviour mathematically. We then determine any 'unknown' coefficients by 'fitting' it to the available data.
We use a particular class of more generic model; for example: a linear regression model, or a standard stochastic time-series model. We then choose such a model, and again fit it to the available data.

In general, when we model complex problems, we often need to make simplifying assumptions in order to make use of standard methodologies. In many cases we (safely) assume linearity, and use linear models and analysis methods.

Unfortunately, there are many problems that are fundamentally non-linear, and where linear approximations break down quite quickly. There are also situations where we might believe that a system can, in principle, be modelled, but we do not have a sufficiently clear insight into its internal dynamics to be able to construct such a model.

In addition, if we use a complex model for a system, fit this to a limited data-set, and then try to use it to predict or extrapolate that system's behaviour - we may not be able to trust the results that it produces.

Genetic Algorithms - Evolving Solutions by Virtual Natural Selection

GAVONTS uses an advanced Genetic Algorithm - which is a particularly powerful and robust optimisation technique, and can be used to address a wide range of problems. Using this approach, we apply the same principles of natural selection that drive biological evolution in the natural world, to similarly drive the solution of complex mathematical problems in a virtual world.

In GAVONTS, the process of evolution occurs by simulating the passage of many generations of a population of individual 'solution functions'. For each generation, some of the worst performers are removed, and replaced with new ones that are created by 'breeding' and 'mutating' individuals from within the surviving population.

How 'good' or 'bad' each individual solution is, is determined according to an error metric that is specific to the particular problem being tackled. However, applying GAs in practice is rarely straightforward, the main difficulties that tend to be encountered, are:

Their efficient progress relies on the ability to 'encode' individual solutions, so that both 'breeding' and 'mutation' of individuals can produces useful results sufficiently often.
These algorithms can be extremely computationally intensive, even for relatively modest problems.

We deal with these issues, firstly, through the use of a proprietary GA modelling framework, with a number of built-in features designed to improve its performance; and secondly by using our xml-space parallel processing techniques to handle the computational load.

The result is that the GAVONTS meta-model framework can create and refine solutions with little or no human guidance, and can solve complex problems in ways that can be both surprising and effective.