C Clack

Selective Crossover in Genetic Algorithms: An Empirical Study

The performance of a genetic algorithm (GA) is dependent on many factors: the type of crossover operator, the rate of crossover, the rate of mutation, population size, and the encoding used are just a few examples. Currently, GA practitioners pick and choose GA parameters empirically until they achieve adequate performance for a given problem. In this paper we have isolated one such parameter: the crossover operator. The motivation for this study is to provide an adaptive crossover operator that gives best overall performance on a large set of problems. A new adaptive crossover operator “selective crossover” is proposed and is compared with two-point and uniform crossover on a problem generator where epistasis can be varied and on trap functions where deception can be varied. We provide empirical results which show that selective crossover is more efficient than two-point and uniform crossover across a representative set of search problems containing epistasis.

Autonomous document classification for business

With the continuing exponential growth of the Internet and the more recent growth of business Intranets, the commercial world is becoming increasingly aware of the problem of electronic information overload. This has encouraged interest in developing agents/softbots that can act as electronic personal assistants and can develop and adapt representations of users information needs, commonly known as profiles. As the result of collaborative research with Friends of the Earth, a leading environmental campaigning organization, we have developed a general purpose information classification agent architecture and are applying it to the problem of document classification and routing. Collaboration with Friends of the Earth allows us to test our ideas in a non-academic context involving high volumes of documents. We use the technique of genetic programming (GP), (Koza & Rice 1992), to evolve classifying agents. This is a novel approach for document classification, where each agent evolves a parse-tree representation of a users particular information need. The other unusual features of our research are the longevity of our agents and the fact that they undergo a continual training process; feedback from the user enables the agent to adapt to the users long-term information requirements.

Evolving robust GP solutions for hedge fund stock selection in emerging markets

Stock selection for hedge fund portfolios is a challenging problem for Genetic Programming (GP) because the markets (the environment in which the GP solution must survive) are dynamic, unpredictable and unforgiving. How can GP be improved so that solutions are produced that are robust to non-trivial changes in the environment? We explore an approach that uses subsets of extreme environments during training.

GRIP - A high-performance architecture for parallel graph reduction

GRIP is a high-performance parallel machine designed to execute functional programs using supercombinator graph reduction. It uses a high-bandwidth bus to provide access to a large, distributed shared memory, using intelligent memory units and packet-switching protocols to increase the number of processors which the bus can support. GRIP is also being programmed to support parallel Prolog and DACTL.

Performance-Enhanced Genetic Programming

Genetic Programming is increasing in popularity as the basis for a wide range of learning algorithms. However, the technique has to date only been successfully applied to modest tasks because of the performance overheads of evolving a large number of data structures, many of which do not correspond to a valid program. We address this problem directly and demonstrate how the evolutionary process can be achieved with much greater efficiency through the use of a formally-based representation and strong typing. We report initial experimental results which demonstrate that our technique exhibits significantly better performance than previous work.

Strictness analysis—a practical approach

Significant improvements in performance arise if we can arrange for parallel execution of programs. The absence of side effects in functional languages allows concurrent evaluation of the program, but in lazy implementations this risks wasting work by evaluating expressions which are subsequently discarded. We discuss the use of strictness analysis to determine at compile time which parts of program evaluation can safely be carried out concurrently. We give a practical explanation of this technique, concentrating particularly on the problem of finding fixed points.

Behavioural GP diversity for dynamic environments: an application in hedge fund investment

We present a new mechanism for preserving phenotypic behavioural diversity in a Genetic Programming application for hedge fund portfolio optimization, and provide experimental results on real-world data that indicate the importance of phenotypic behavioural diversity both in achieving higher fitness and in improving the adaptability of the GP population for continuous learning.

Lexical profiling: theory and practice

This paper addresses the issue of analysing the run-time behaviour of lazy, higher-order functional programs. We examine the difference between the way that functional programmers and functional language implementors view program behaviour. Existing profiling techniques are discussed and a new technique is proposed which produces results that are straightforward for programmers to assimilate. The new technique, which we call lexical profiling, collects information about the run-time behaviour of functional programs, and reports the results with respect to the original source code rather than simply listing the actions performed at run-time. Lexical profiling complements implementation-specific profiling and is important because it provides a view of program activity which is largely independent of the underlying evaluation mechanism. Using the lexical profiler, programmers may easily relate results back to the source program. We give a full implementation of the lexical profiling technique for a sequential, interpretive graph reduction engine, and extensions for compiled and parallel graph reduction are discussed.

A formalism for multi-level emergent behaviours in designed component-based systems and agent-based simulations

There currently exists no means of specifying or analysing specific emergent behaviours in designed multi-component systems. For this reason, important questions about the lower level mechanisms giving rise to emergent behaviours cannot be resolved.

Identifying Multi-Level Emergent Behaviors in Agent-Directed Simulations using Complex Event Type Specifications

Agent-directed simulations (ADS) are used in many domains to study complex systems. These are systems where non-linear effects can result from these emergent behaviors, making them difficult to analyze and predict. Correspondingly, in ADS, as well as explicitly specified behaviors of individual agents, higher level behaviors can emerge spontaneously from agent action sequences and agent—agent interactions. We have previously introduced the complex event formalism for specifying emergent behaviors in dynamically executing ADS [1, 2]. Based on the formalism, we also described a method for detecting and analyzing emergent behaviors in multi-agent simulations, giving us an effective means of studying, and a more reliably way of predicting, these systems. Complex event types define sets of multi-dimensional structures of interrelated events arising from the actions of one or more agents. They are therefore directly related to the agent specifications, which determine the behavior of individual agents. Although the abstract constructs of the formalism have already been introduced in [1] and [2], they have not yet been related to a specific agent-based specification language. Here, we define the constructs in terms of the X-machine formalism, which is widely used to specify multi-agent systems. This extends the existing X-machine framework to model higher level emergent behaviors as well as agent-level state transitions. Thus, emergent behaviors at any level of abstraction can be specified for detection and analysis in a dynamically executing ADS.