Edmund M. A. Ronald

Design, observation, surprise! A test of emergence

The field of artificial life (Alife) is replete with documented instances of emergence, though debate still persists as to the meaning of this term. We contend that, in the absence of an acceptable definition, researchers in the field would be well served by adopting an emergence certification mark that would garner approval from the Alife community. Toward this end, we propose an emergence test, namely, criteria by which one can justify conferring the emergence label.

Neuro-genetic truck backer-upper controller

The precise docking of a truck at a loading dock has been proposed in (Nguyen and Widrow, 1990) as a benchmark problem for non-linear control by neural-nets. The main difficulty is that backpropagation is not a priori suitable as a learning paradigm, because no set of training vectors is available: It is non-trivial to find solution trajectories that dock the truck from anywhere in the loading yard. In this paper we show how a genetic algorithm can evolve the weights of a feedforward 3-layer neural net that solves the control problem for a given starting state, achieving a short trajectory from starting point to goal. The fitness of a net in the population is a function of both the nearest position from the goal and the distance travelled. The influence of input data renormalisation on trajectory precision is also discussed.<<ETX>>

Testing for Emergence in Artificial Life

The field of artificial life (Alife) is replete with documented instances of emergence, though debate still persists as to the meaning of this term. In the absence of a formal definition, researchers in the field would be well served by adopting an emergence certification mark which would garner approval from the Alife community. We propose an emergence test, consisting of three criteria--design, observation, and surprise--for conferring the emergence label.

Surprise versus Unsurprise: Implications of Emergence in Robotics

Reference EPFL-ARTICLE-28723doi:10.1016/S0921-8890(01)00149-XView record in Web of Science Record created on 2004-11-30, modified on 2016-08-08

A Simple Cellular Automaton that Solves the Density and Ordering Problems

We show that there exists a simple solution to the density problem in cellular automata, under fixed boundary conditions, in contrast to previously used periodic ones.

Genetic Lander: An Experiment in Accurate Neuro-Genetic Control

The control problem of soft-landing a toy lunar module simulation is investigated in the context of neural nets. While traditional supervised back-propagation training is inappropriate for lack of training exemplars, genetic algorithms allow a controller to be evolved without difficulty: Evolution is a form of unsupervised learning. A novelty introduced in this paper is the presentation of additional renormalized inputs to the net; experiments indicate that the presence of such inputs allows precision of control to be attained faster, when learning time is measured by the number of generations for which the GA must run to attain a certain mean performance.

Million module neural systems evolution

This position paper discusses the evolution of multi-module neural net systems, where the number of neural net modules is up to ten million (i.e. an “artificial brain”). ATRs “CAM-Brain” Project [de Garis 1993, 1996] has progressed to the point where it is technically possible (using a new FPGA (Field Programmable Gate Array) based evolvable hardware (EHW or E-Hard) system to be completed by the spring of 1998 [Korkin & de Garis 1997]) to begin to evolve and build an artificial brain containing 10,000 neural net modules. This development raises the prospect that within a few years these numbers will rapidly increase. This paper introduces some issues that such massive system-building will generate. The immediate question is “What should we evolve?” This paper presents some suggested evolvable system targets containing N neural net modules, where N = 100; 1000; 10,000; 100,000; 1,000,000; 10,000,000 with an emphasis on the N = 100 case, for purposes of illustration. The issues involved are not only of a conceptual and evolutionary engineering nature, but (when N is large) economic, managerial and even political as well.

How Long Does It Take to Evolve a Neural Net

This paper deals with technical issues relevant to artificial neural net (ANN) training by genetic algorithms. Neural nets have applications ranging from perception to control; in the context of control, achieving great precision is more critical than in pattern recognition or classification tasks. In previous work, the authors have found that when employing genetic search to train a net, both precision and training speed can be greatly enhanced by an input renormalization technique. In this paper we investigate the automatic tuning of such renormalization coefficients, as well as the tuning of the slopes of the transfer functions of the individual neurons in the net. Waiting time analysis is presented as an alternative to the classical ”mean performance” interpretation of GA experiments. It is felt that it provides a more realistic evaluation of the real-world usefulness of a GA.