On dynamical genetic programming: simple Boolean networks in learning classifier systems

Abstract

Abstract. Many representations have been presented to enable the effective evolution of computer programs. Turing was perhaps the first to present a general scheme by which to achieve this end.Significantly, Turing proposed a form of discrete dynamical system and yet dynamical representations remain almost unexplored within conventional genetic programming. This paper presents results from an initial investigation into using simple dynamical genetic programming representations within a Learning Classifier System. It is shown possible to evolve ensembles of dynamical Boolean function networks to solve versions of the well-known multiplexer problem. Both synchronous and asynchronous systems are considered.

Evolving Boolean Networks on Tunable Fitness Landscapes

“…This general result has long been predicted by Kauffman [34] and is now supported by data from real GRN which appear to be relatively sparsely connected: on average it seems 1.5 ≤ B ≤ 2 (see [39] for discussions). The aforementioned work evolving RBN for machine learning problems found a B value of around 2 typically emerged also [11]. This result can be seen to extend that previously reported by [48] who also found longer attractors harder to evolve.…”

“…In the above, fitness is calculated from the state of the N trait nodes on the step after T network update cycles, i.e., within an attractor (see [11]). A number of examples of related previous work discussed in Section IV have sought to evolve temporal behavior, i.e., particular sequences of gene activity wherein the action of the GRN is sampled on every update cycle, i.e., up to and potentially including within an attractor.…”

“…Sipper and Ruppin [52] evolved RBN for the well-known density task. Bull [11] evolved RBN ensembles to solve simple machine learning problems.…”

Arithmetic dynamical genetic programming in the XCSF Learning Classifier System

“…A number of representations have been previously presented beyond this scheme however, including real numbers, fuzzy logic, genetic programming, and artificial neural networks. Recently, we [2] [3] investigated the use of a Dynamical Genetic Programming representation scheme (DGP) [4] within Learning Classifier Systems. It was shown possible to evolve ensembles of Random Boolean Networks (RBN) [5] as a discrete, temporally dynamic, knowledge representation scheme within LCS.…”

On the evolution of Boolean networks for computation: a guide RNA mechanism

“…The same approach has been used to explore attractor stability [15] and to model real regulatory network data, eg, see [16] for an example using probabilistic RBN. Sipper and Ruppin [17] evolved RBN for the well-known density task and Bull [18] has evolved RBN ensembles to solve benchmark machine learning problems. RBN appear to have useful properties for computation, such as fault tolerance (eg, see [19]).…”