Sonia Schulenburg

Hyper-Heuristics: An Emerging Direction in Modern Search Technology

This chapter introduces and overviews an emerging methodology in search and optimisation. One of the key aims of these new approaches, which have been termed hyperheuristics, is to raise the level of generality at which optimisation systems can operate. An objective is that hyper-heuristics will lead to more general systems that are able to handle a wide range of problem domains rather than current meta-heuristic technology which tends to be customised to a particular problem or a narrow class of problems. Hyper-heuristics are broadly concerned with intelligently choosing the right heuristic or algorithm in a given situation. Of course, a hyper-heuristic can be (often is) a (meta-)heuristic and it can operate on (meta-)heuristics. In a certain sense, a hyper-heuristic works at a higher level when compared with the typical application of meta-heuristics to optimisation problems, i.e., a hyper-heuristic could be thought of as a (meta)-heuristic which operates on lower level (meta-)heuristics. In this chapter we will introduce the idea and give a brief history of this emerging area. In addition, we will review some of the latest work to be published in the field.

Learning a procedure that can solve hard bin-packing problems: a new GA-based approach to hyper-heuristics

The idea underlying hyper-heuristics is to discover some combination of familiar, straightforward heuristics that performs very well across a whole range of problems. To be worthwhile, such a combination should outperform all of the constituent heuristics. In this paper we describe a novel messy-GA-based approach that learns such a heuristic combination for solving one-dimensional bin-packing problems. When applied to a large set of benchmark problems, the learned procedure finds an optimal solution for nearly 80% of them, and for the rest produces an answer very close to optimal. When compared with its own constituent heuristics, it ranks first in 98% of the problems.

Explorations in LCS Models of Stock Trading

In previous papers we have described the basic elements for building an economic model consisting of a group of artificial traders functioning and adapting in an environment containing real stock market information. We have analysed the feasibility of the proposed approach by comparing the final wealth generated by such agents over a period of time, against the wealth of a number of well known investment strategies, including the bank, buy-and-hold and trend-following strategies. In this paper we review classical economic theories and introduce a new strategy inspired by the Efficient Market Hypothesis (named here random walk to compare the performance of our traders. In order to build better trader models we must increase our understanding about how artificial agents learn and develop; in this paper we address a number of design issues, including the analysis of information sets and evolved strategies. Specifically, the results presented here correspond to the stock of IBM.

Strength and Money: An LCS Approach to Increasing Returns

This paper reports on a number of experiments where three different groups of artificial agents learn, forecast and trade their holdings in a real stock market scenario given exogeneously in the form of easily-obtained stock statistics such as various price moving averages, first difference in prices, volume ratios, etc. These artificial agent-types trade while learning during - in most cases - a ten year period. They normally start at the beginning of the year 1990 with a fixed initial wealth to trade over two assets (a bond and a stock) and end in the second half of the year 2000. The adaptive agents are represented as Learning Classifier Systems (LCSs), that is, as sets of bit-encoded rules. Each condition bit expresses the truth or falsehood of a certain real market condition. The actual conditions used differ between agents. The forecasting performance is then compared against the performance of the buy-and-hold strategy, a trend-following strategy and finally against the bank investment over the same period of time at a fixed compound interest rate. To make the experiments as real as possible, agents pay commissions on every trade. The results so far suggest that this is an excellent approach to make trading decisions in the stock market.

Collective behavior based hierarchical XCS

This paper attempts to extend the XCS research by analyzing the impact of information exchange between XCS agents on classifier performance. Two types of information are exchanged and combined to improve classification performance. The first uncovers information contained in the signal patterns of collections of Homogeneous XCS classifiers. This information is used to determine which subsets of the state-space the XCS can be expected to be accurately classified. The second combines the results of XCS agents that are each tasked to solve different portions of the original problem. Results on the multiplexer (6, 11) indicate that given accurate problem domain assumptions, the Collective Behavior (CB-HXCS) method shows promise. Results show - at least in simulated multiplexer environments - that the HXCS is able to solve a well defined problem with less data than an individual XCS. This approach seems very promissing in real-world applications where data is incomplete, expensive or unreliable such as in financial or medical domains.