Denis Robilliard

Population parallel GP on the G80 GPU

The availability of low cost powerful parallel graphics cards has stimulated a trend to port GP on Graphics Processing Units (GPUs). Previous works on GPUs have shown evaluation phase speedups for large training cases sets. Using the CUDA language on the G80 GPU, we show it is possible to efficiently interpret several GP programs in parallel, thus obtaining speedups also for small training sets starting at less than 100 training cases. Our scheme was embedded in the well-known ECJ library, providing an easy entry point for owners of G80 GPUs.

Genetic programming on graphics processing units

The availability of low cost powerful parallel graphics cards has stimulated the port of Genetic Programming (GP) on Graphics Processing Units (GPUs). Our work focuses on the possibilities offered by Nvidia G80 GPUs when programmed in the CUDA language. In a first work we have showed that this setup allows to develop fine grain parallelization schemes to evaluate several GP programs in parallel, while obtaining speedups for usual training sets and program sizes. Here we present another parallelization scheme and optimizations about program representation and use of GPU fast memory. This increases the computation speed about three times faster, up to 4 billion GP operations per second. The code has been developed within the well known ECJ library and is open source.

High performance genetic programming on GPU

The availability of low cost powerful parallel graphics cards has stimulated the port of Genetic Programming (GP) on Graphics Processing Units (GPUs). Our work focuses on the possibilities offered by Nvidia G80 GPUs when programmed in the CUDA language. We compare two parallelization schemes that evaluate several GP programs in parallel. We show that the fine grain distribution of computations over the elementary processors greatly impacts performances. We also present memory and representation optimizations that further enhance computation speed, up to 2.8 billion GP operations per second. The code has been developed with the well known ECJ library.

Inversion of oceanic constituents in case I and II waters with genetic programming algorithms.

A stochastic inverse technique based on agenetic programming (GP) algorithm was developed toinvert oceanic constituents from simulated data for case I and case II water applications. The simulations were carried out with the Ordre Successifs Ocean Atmosphere (OSOA) radiative transfer model. They include the effects of oceanic substances such as algal-related chlorophyll, nonchlorophyllous suspended matter, and dissolved organic matter. The synthetic data set also takes into account the directional effects of particles through a variation of their phase function that makes the simulated data realistic. It is shown that GP can be successfully applied to the inverse problem with acceptable stability in the presence of realistic noise in the data. GP is compared with neural network methodology for case I waters; GP exhibits similar retrieval accuracy, which is greater than for traditional techniques such as band ratio algorithms. The application of GP to real satellite data [a Sea-viewing Wide Field-of-view Sensor (SeaWiFS)] was also carried out for case I waters as a validation. Good agreement was obtained when GP results were compared with the SeaWiFS empirical algorithm. For case II waters the accuracy of GP is less than 33%, which remains satisfactory, at the present time, for remote-sensing purposes.

Tarpeian bloat control and generalization accuracy

In this paper we focus on machine-learning issues solved with Genetic Programming (GP). Excessive code growth or bloat often happens in GP, greatly slowing down the evolution process. In Pol03, Poli proposed the Tarpeian Control method to reduce bloat, but possible side-effects of this method on the generalization accuracy of GP hypotheses remained to be tested. In particular, since Tarpeian Control puts a brake on code growth, it could behave as a kind of Occams razor, promoting shorter hypotheses more able to extend their knowledge to cases apart from any learning steps. To answer this question, we experiment Tarpeian Control with symbolic regression. The results are contrasted, showing that it can either increase or reduce the generalization power of GP hypotheses, depending on the problem at hand. Experiments also confirm the decrease in size of programs. We conclude that Tarpeian Control might be useful if carefully tuned to the problem at hand.

Exploring Overfitting in Genetic Programming

The problem of overfitting (focusing closely on examples at the loss of generalization power) is encountered in all supervised machine learning schemes. This study is dedicated to explore some aspects of overfitting in the particular case of genetic programming. After recalling the causes usually invoked to explain overfitting such as hypothesis complexity or noisy learning examples, we test and compare the resistance to overfitting on three variants of genetic programming algorithms (basic GP, sizefair crossover GP and GP with boosting) on two benchmarks, a symbolic regression and a classification problem. We propose guidelines based on these results to help reduce overfitting with genetic programming.

A Bit-Wise Epistasis Measure for Binary Search Spaces

The epistatic variance has been introduced by Davidor as a tool for the evaluation of interdependences between genes, thus possibly giving clues about the difficulty of optimizing functions with genetic algorithms (GAs). Despite its theoretical grounding in Walsh function analysis, several studies have shown its weakness as a predictor of GAs results. In this paper, we focus on binary search spaces and propose to measure epistatic effect on the level of individual genes, an approach that we call bit-wise epistasis. We give examples of this measure on several well-known test problems, then we take into account this supplementary information to improve the performances of evolutionary algorithms. We conclude by pointing towards possible extensions of this concept to real size problems.

Applying Boosting Techniques to Genetic Programming

This article deals with an improvement for genetic programming based on a technique originating from the machine learning field: boosting. In a first part of this paper, we test the improvements offered by boosting on binary problems. Then we propose to deal with regression problems, and propose an algorithm, called GPboost, that keeps closer to the original idea of distribution in Adaboost than what has been done in previous implementation of boosting for genetic programming.

Santa fe trail hazards

This paper focuses on methodological problems associated to the famous Santa Fe Trail (SFT) problem, a very common benchmark for evaluating Genetic Programming (GP) algorithms, introduced by Koza in its first book on GP. We put in evidence the difficulty to ensure fair comparisons especially with new genotype representations as found in works on grammar-based automatic programming, such as Grammatical Evolution, and Bayesian Automatic Programming. We extend a work by Langdon et al. by measuring the effort to solve SFT by random search with different time steps limits and a reduced but semantically equivalent function set.

Investigating monte-carlo methods on the weak schur problem

Nested Monte-Carlo Search (NMC) and Nested Rollout Policy Adaptation (NRPA) are Monte-Carlo tree search algorithms that have proved their efficiency at solving one-player game problems, such as morpion solitaire or sudoku 16x16, showing that these heuristics could potentially be applied to constraint problems. In the field of Ramsey theory, the weak Schur numberWS(k) is the largest integer n for which their exists a partition into k subsets of the integers [1,n] such that there is no x<y<z all in the same subset with x+y=z. Several studies have tackled the search for better lower bounds for the Weak Schur numbers WS(k), k≥4. In this paper we investigate this problem using NMC and NRPA, and obtain a new lower bound for WS(6), namely WS(6)≥582.