|
  |
|
|
Krasnow Institute > Monday Seminars > Abstracts Synergistic Learning in an Evolving Population of Agents: A Computational Model of Cultural Coevolution Larry Hunter Abstract: In this talk, I will describe preliminary work on an "evolutionary" machine learning system which uses two independent but interacting inheritance systems, inspired by William Durham's mathematical models of human genetic and cultural evolution. The system consists of an evolving population of several different types of machine learning programs, all trying to solve the same supervised classification task. The two independent inheritance systems involve (a) the free parameters of the learning methods and (b) the input representations used by the learning methods. Each learner requires a vector of free parameters (e.g. depth of lookahead, pruning method, number of hidden nodes) which are taken to constitute its 'genome.' These parameters are adjusted via a genetic algorithm. Each program also requires a set of transformations of given data primitives which constitute its input representation. These transformations are typically boolean or arithmetic combinations of the primitive features. Each transformation maps a combination of data primitives to a single input "feature", so the input representation for the learner is determined by the set of transformations it uses. Each transformation is called a 'meme,' and the collection of transformations used by a learner is its 'memome.' New memes are constructed by analyzing the classifier output by each learner (e.g., by constructive induction, or neural network rule extraction), and the pool of old and new memes are transmitted to learners in the next generation with probability proportional to their fitness. The coevolution process involves executing each learner in the population, using its specific parameter values and input representation. After execution, the output of each learner is analyzed to identify new memes (feature combinations that were useful to that learner) and to determine its cross-validation accuracy and execution time. The fitness of a parameter set is a function of the execution time and cross-validation accuracy of the resulting learner. The fitness of a meme is a function of how many learners used it in their classifiers, how important it was to the performance of the classifier, and how well the classifiers it was part of performed on the task. Variability in the genomes arises from mutation and crossover among the vectors. Variability in the memomes arises from the construction of new memes by individual learners. I will analyze this system from several perspectives: as a practical problem solving tool, as a way of addressing certain fundamental issues in machine learning, and as a model of human cultural coevolution. I will present several suggestive, although preliminary, results regarding effective multistrategy learning, gene/meme interactions, population diversity, and efficient parallel implementation.
The Krasnow Institute for Advanced Study |
|
