Learning Classifier Systems (LCS) constitute a fascinating concept at the intersection of machine learning and evolutionary computation. LCS’s genetic search, generally in combination with reinforcement learning techniques, can be applied to both temporal and spatial problem-solving and promotes powerful search in a wide variety of domains. The LCS concept allows many representations of the learned knowledge from simple production rules to artificial neural networks to linear approximations often in a human readable form.
The concepts underlying LCS have been developed for over 30 years, with the annual International Workshop on Learning Classifier Systems supporting the field since 1992. From 1999 onwards the workshop has been held yearly, in conjunction with PPSN in 2000 and 2002 and with GECCO in 1999, 2001, and from 2003 onwards. This book is the continuation of the six volumes containing selected and revised papers from the previous workshops, published by Springer as LNAI 1813, LNAI 1996, LNAI 2321, LNAI 2661, LNCS 4399, and LNAI 4998.
The articles in this book have been loosely organized into four overlapping themes. Firstly, the breadth of research into LCS and related areas is demonstrated.
Then the ability to approximate complex multidimensional function surfaces is shown by the latest research on computed predictions and piecewise approximations. This work leads on to LCS for complex domains, such as temporal decision-making and continuous domains, whereas traditional learning approaches often require problem-dependent manual tuning of the algorithms and discretization of problem spaces, resulting in a loss of information. Finally, diverse application examples are presented to demonstrate the versatility and broad applicability of the LCS approach.
Pier Luca Lanzi and Daniele Loiacono investigate the use of general-purpose Graphical Processing Units (GPUs), which are becoming increasingly common in evolutionary computation, for speeding up matching of environmental states to rules in LCS. Depending on the problem investigated and representation scheme used, they find that the use of GPUs improves the matching speed by 3 to 50 times when compared with matching with standard CPUs. Association rule mining, where interesting associations in the occurrence of items in streams of unlabelled examples are to be extracted, is addressed by Albert Orriols-Puig and Jorge Casillas. Their novel CSar Michigan-style learning classifier system shows promising results when compared with the benchmark approach to this problem. Stewart Wilson shows that there is still much scope in generating novel approaches with the LCS concept. He proposes an automatic system for creating pattern generators and recognizers based on a three-cornered competitive co-evolutionary algorithm approach.