Alexandros Agapitos

Evolving controllers for simulated car racing using object oriented genetic programming

Several different controller representations are compared on anon-trivial problem in simulated car racing, with respect tolearning speed and final fitness. The controller representations arebased either on Neural Networks or Genetic Programming, and alsodiffer in regards to whether they allow for stateful controllers orjust reactive ones. Evolved GP trees are analysed, and attempts aremade at explaining the performance differences observed.

Learning recursive functions with object oriented genetic programming

This paper describes the evolution of recursive functions within an Object-Oriented Genetic Programming (OOGP) system. We evolved general solutions to factorial, Fibonacci, exponentiation, even-n-Parity, and nth-3. We report the computational effort required to evolve these methods and provide a comparison between crossover and mutation variation operators, and also undirected random search. We found that the evolutionary algorithms performed much better than undirected random search, and thats mutation outperformed crossover on most problems.

Evolving Efficient Recursive Sorting Algorithms

Object Oriented Genetic Programming (OOGP) is applied to the task of evolving general recursive sorting algorithms. We studied the effects of language primitives and fitness functions on the success of the evolutionary process. For language primitives, these were the methods of a simple list processing package. Five different fitness functions based on sequence disorder were evaluated. The time complexity of the successfully evolved algorithms was measured experimentally in terms of the number of method invocations made, and for the best evolved individuals this was best approximated as O(n times log(n)). This is the first time that sorting algorithms of this complexity have been evolved.

Generating diverse opponents with multiobjective evolution

For computational intelligence to be useful in creating game agent AI, we need to focus on creating interesting and believable agents rather than just learn to play the games well. To this end, we propose a way to use multiobjective evolutionary algorithms to automatically create populations of non-player characters (NPCs), such as opponents and collaborators, that are interestingly diverse in behaviour space. Experiments are presented where a number of partially conflicting objectives are defined for racing game competitors, and multiobjective evolution of Genetic Programming-based controllers yield pareto fronts of interesting controllers.

Evolving modular recursive sorting algorithms

A fundamental issue in evolutionary learning is the definition of the solution representation language. We present the application of Object Oriented Genetic Programming to the task of coevolving general recursive sorting algorithms along with their primitive representation alphabet. We report the computational effort required to evolve target solutions and provide a comparison between crossover and mutation variation operators, and also undirected random search. We found that the induction of evolved method signatures (typed parameters and return type) can be realized through an evolutionary fitness-driven process. We also found that the evolutionary algorithm outperformed undirected random search, and that mutation performed better than crossover in this problem domain. The main result is that modular sorting algorithms can be evolved.

An investigation of fitness sharing with semantic and syntactic distance metrics

This paper investigates the efficiency of using semantic and syntactic distance metrics in fitness sharing with Genetic Programming (GP). We modify the implementation of fitness sharing to speed up its execution, and used two distance metrics in calculating the distance between individuals in fitness sharing: semantic distance and syntactic distance. We applied fitness sharing with these two distance metrics to a class of real-valued symbolic regression. Experimental results show that using semantic distance in fitness sharing helps to significantly improve the performance of GP more frequently, and results in faster execution times than with the syntactic distance. Moreover, we also analyse the impact of the fitness sharing parameters on GP performance helping to indicate appropriate values for fitness sharing using a semantic distance metric.

Ubiquitous robotics in physical human action recognition: A comparison between dynamic ANNs and GP

Two different classifier representations based on dynamic artificial neural networks (ANNs) and genetic programming (GP) are being compared on a human action recognition task by an ubiquitous mobile robot. The classification methodologies used, process time series generated by an indoor ubiquitous 3D tracker which generates spatial points based on 23 reflectable markers attached on a human body. This investigation focuses mainly on class discrimination of normal and aggressive action recognition performed by an architecture which implements an interconnection between an ubiquitous 3D sensory tracker system and a mobile robot to perceive, process, and classify physical human actions. The 3D tracker and the robot are used as a perception-to-action architecture to process physical activities generated by human subjects. Both classifiers process the activity time series to eventually generate surveillance assessment reports by generating evaluation statistics indicating the classification accuracy of the actions recognized.

A preliminary investigation of overfitting in evolutionary driven model induction: implications for financial modelling

This paper investigates the effects of early stopping as a method to counteract overfitting in evolutionary data modelling using Genetic Programming. Early stopping has been proposed as a method to avoid model overtraining, which has been shown to lead to a significant degradation of out-of-sample performance. If we assume some sort of performance metric maximisation, the most widely used early training stopping criterion is the moment within the learning process that an unbiased estimate of the performance of the model begins to decrease after a strictly monotonic increase through the earlier learning iterations. We are conducting an initial investigation on the effects of early stopping in the performance of Genetic Programming in symbolic regression and financial modelling. Empirical results suggest that early stopping using the above criterion increases the extrapolation abilities of symbolic regression models, but is by no means the optimal training-stopping criterion in the case of a real-world financial dataset.

On the genetic programming of time-series predictors for supply chain management

Single and multi-step time-series predictors were evolved for forecasting minimum bidding prices in a simulated supply chain management scenario. Evolved programs were allowed to use primitives that facilitate the statistical analysis of historical data. An investigation of the relationships between the use of such primitives and the induction of both accurate and predictive solutions was made, with the statistics calculated based on three input data transformation methods: integral, differential, and rational. Results are presented showing which features work best for both single-step and multi-step predictions.

Evolutionary learning of technical trading rules without data-mining bias

In this paper we investigate the profitability of evolved technical trading rules when controlling for data-mining bias. For the first time in the evolutionary computation literature, a comprehensive test for a rules statistical significance using Hansens Superior Predictive Ability is explicitly taken into account in the fitness function, and multi-objective evolutionary optimisation is employed to drive the search towards individual rules with better generalisation abilities. Empirical results on a spot foreign-exchange market index suggest that increased out-of-sample performance can be obtained after accounting for data-mining bias effects in a multi-objective fitness function, as compared to a single-criterion fitness measure that considers solely the average return.