Search

The idea that lifetime learning can have a significant effect on life history evolution has recently been explored using a series of artificial life simulations. These involved populations of competing individuals evolving by natural selection to learn to perform well on simplified abstract tasks, with the learning consisting of identifying regularities in their environment. In reality, there is more to learning than that type of direct individual experience, because it often includes a substantial degree of social learning that involves various forms of imitation of what other individuals have learned before them. This article rectifies that omission by incorporating memes and imitative learning into revised versions of the previous approach. To do this reliably requires formulating and testing a general framework for meme-based simulations that will enable more complete investigations of learning as a factor in any life history evolution scenarios. It does that by simulating imitative information transfer in terms of memes being passed between individuals, and developing a process for merging that information with the (possibly inconsistent) information acquired by direct experience, leading to a consistent overall body of learning. The proposed framework is tested on a range of learning variations and a representative set of life history factors to confirm the robustness of the approach. The simulations presented illustrate the types of interactions and tradeoffs that can emerge, and indicate the kinds of species-specific models that could be developed with this approach in the future.

Display Omitted A multiobjective evolutionary algorithm for vehicle routing problems with backhauls.Our approach equals or improves upon some of the best-known single-objective results.Instances have few bi-objective conflicts; SSMOEA, NSGA-II and MOEA/D perform well.Large tri-objective solution sets and significant differences in both algorithms.Triobjective setting finds comparable solutions to those from bi-objective settings. The vehicle routing problem (VRP) is an important aspect of transportation logistics with many variants. This paper studies the VRP with backhauls (VRPB) in which the set of customers is partitioned into two subsets: linehaul customers requiring a quantity of product to be delivered, and backhaul customers with a quantity to be picked up. The basic VRPB involves finding a collection of routes with minimum cost, such that all linehaul and backhaul customers are serviced. A common variant is the VRP with selective backhauls (VRPSB), where the collection from backhaul customers is optional. For most real world applications, the number of vehicles, the total travel cost, and the uncollected backhauls are all important objectives to be minimized, so the VRPB needs to be tackled as a multi-objective problem. In this paper, a similarity-based selection evolutionary algorithm approach is proposed for finding improved multi-objective solutions for VRPB, VRPSB, and two further generalizations of them, with fully multi-objective performance evaluation.

The Artificial Bee Colony (ABC) is a swarm intelligence algorithm for optimization that has previously been applied to the training of neural networks. This paper examines more carefully the performance of the ABC algo- rithm for optimizing the connection weights of feed-forward neural networks for classification tasks, and presents a more rigorous comparison with the traditional Back-Propagation (BP) training algorithm. The empirical results for bench- mark problems demonstrate that using the standard "stop- ping early" approach with optimized learning parameters leads to improved BP performance over the previous com- parative study, and that a simple variation of the ABC approach provides improved ABC performance too. With bothimprovementsapplied,theABCapproachdoesperform very well on small problems, but the generalization perfor- mances achieved are only significantly better than standard BP on one out of six datasets, and the training times increase rapidly as the size of the problem grows. If different, evo- lutionary optimized, BP learning rates are allowed for the two layers of the neural network, BP is significantly better than the ABC on two of the six datasets, and not significantly different on the other four.

In a previous article, we presented a systematic computational study of the extraction of semantic representations from the word-word co-occurrence statistics of large text corpora. The conclusion was that semantic vectors of pointwise mutual information values from very small co-occurrence windows, together with a cosine distance measure, consistently resulted in the best representations across a range of psychologically relevant semantic tasks. This article extends that study by investigating the use of three further factors--namely, the application of stop-lists, word stemming, and dimensionality reduction using singular value decomposition (SVD)--that have been used to provide improved performance elsewhere. It also introduces an additional semantic task and explores the advantages of using a much larger corpus. This leads to the discovery and analysis of improved SVD-based methods for generating semantic representations (that provide new state-of-the-art performance on a standard TOEFL task) and the identification and discussion of problems and misleading results that can arise without a full systematic study.

An artificial life approach is taken to explore the effect that lifetime learning can have on the evolution of certain life history traits, in particular the periods of protection that parents offer their young, and the age at first reproduction of those young. The study begins by simulating the evolution of simple artificial neural network systems that must learn quickly to perform well on simple classification tasks, and determining if and when extended periods of parental protection emerge. It is concluded that longer periods of parental protection of children do offer clear learning advantages and better adult performance, but only if procreation is not allowed during the protection period. In this case, a compromise protection period evolves that balances the improved learning performance against reduced procreation period. The crucial properties of the neural learning processes are then abstracted out to explore the possibility of studying the effect of learning more generally and with better computational efficiency. Throughout, the implications of these simulations for more realistic scenarios are discussed.

The idea that at least some aspects of word meaning can be induced from patterns of word co-occurrence is becoming increasingly popular. However, there is less agreement about the precise computations involved, and the appropriate tests to distinguish between the various possibilities. It is important that the effect of the relevant design choices and parameter values are understood if psychological models using these methods are to be reliably evaluated and compared. In this article, we present a systematic exploration of the principal computational possibilities for formulating and validating representations of word meanings from word co-occurrence statistics. We find that, once we have identified the best procedures, a very simple approach is surprisingly successful and robust over a range of psychologically relevant evaluation measures.

Modularity in the human brain remains a controversial issue, with disagreement over the nature of the modules that exist, and why, when, and how they emerge. It is a natural assumption that modularity offers some form of computational advantage, and hence evolution by natural selection has translated those advantages into the kind of modular neural structures familiar to cognitive scientists. However, simulations of the evolution of simplified neural systems have shown that, in many cases, it is actually non-modular architectures that are most efficient. In this paper, the relevant issues are discussed and a series of simulations are presented that reveal crucial dependencies on the details of the learning algorithms and tasks that are being modelled, and the importance of taking into account known physical brain constraints, such as the degree of neural connectivity. A pattern is established which provides one explanation of why modularity should emerge reliably across a range of neural processing tasks.

PurposeThe purpose of this paper is to discuss the application of computational intelligence techniques to the field of industrial robot control.Design/methodology/approachThe core ideas behind using neural computation, evolutionary computation, and fuzzy logic techniques are presented, along with a selection of specific real‐world applications.FindingsTheir practical advantages and disadvantages relative to more traditional approaches are made clear.Originality/valueThe reader will appreciate the power of computational intelligence techniques for industrial robot control, and hopefully be encouraged to explore further the possibility of using them to achieve improved performance in their own applications.

Autonomous neural network systems typically require fast learning and good generalization performance, and there is potentially a trade-off between the two. The use of evolutionary techniques to improve the learning abilities of neural network systems is now widespread. However, there are a range of different evolutionary approaches that could be applied, and no systematic investigation has been carried out to find which work best. In this paper, such an investigation is presented, and it is shown that a range of evolutionary techniques can generate high performance networks, but they often lead to unwanted side effects, such as occasional instances of very poor performance. The nature of these problems are explored further, and it is shown how the evolution of age dependent plasticities and/or the use of ensemble techniques can alleviate them. A range of techniques are thus identified, with differing properties, that can be matched to the specific requirements of each application.

Setting up a neural network with a learning algorithm that determines how it can best operate is an efficient way to formulate control systems for many engineering applications, and is often much more feasible than direct programming. This paper examines three important aspects of this approach: the details of the cost function that is used with the gradient descent learning algorithm, how the resulting system depends on the initial pre-learning connection weights, and how the resulting system depends on the pattern of learning rates chosen for the different components of the system. We explore these issues by explicit simulations of a toy model that is a simplified abstraction of part of the human oculomotor control system. This allows us to compare our system with that produced by human evolution and development. We can then go on to consider how we might improve on the human system and apply what we have learnt to control systems that have no human analogue.

The Artificial Bee Colony (ABC) is a recently introduced swarm intelligence algorithm for optimization, that has previously been applied successfully to the training of neural networks. This paper explores more carefully the performance of the ABC algorithm for optimizing the connection weights of feed-forward neural networks for classification tasks, and presents a more rigorous comparison with the traditional Back-Propagation (BP) training algorithm. The empirical results show that using the standard “stopping early” approach with optimized learning parameters leads to improved BP performance over the previous comparative study, and that a simple variation of the ABC approach provides improved ABC performance too. With both improvements applied, we conclude that the ABC approach does perform very well on small problems, but the generalization performances achieved are only significantly better than standard BP on one out of six datasets, and the training times increase rapidly as the size of the problem grows.

After a brief review of the different types and causes of ambiguous training data and the problems of learning from such data, a class of multi-target models are presented which suggest that neural networks are even better at solving these problems than previously realized. They are able to learn which non-ambiguous subset of a larger ambiguous set of training data best captures any underlying regularities in that data and hence optimize generalization while minimizing the problems of overtraining. It is also shown how the deliberate generation of ambiguous training data can begin to solve some of the long-standing representational problems of mapping time sequences, such as the alignment problem for reading and spelling. The general ideas are illustrated throughout with the well- known problem of tex-to-phoneme conversion, and detailed results of a range of neural network simulations are presented.

Connectionist techniques are increasingly being used to model cognitive function with a view to providing extensions, elaborations or replacements of earlier “box and arrow” models. Networks of simplified processing units loosely based on real neurons are set up with architectures based on known physiology, trained to perform appropriately simplified versions of real tasks, and iteratively refined by checking their performance against humans. Such systems can still be linked together as in the old box and arrow models, with all the old explanations of patient data carrying through, but now we can examine the details of the degradation of the various components, and removing neurons or connections constitute natural analogues of real brain damage. We can also question the validity of the old assumptions of neuropsychological inference, and explore the possibility that processing is actually more distributed and interactive than the older models implied (Bullinaria, 2002). Here I shall outline what I consider to be the essential ideas.

Modularity is known to have benefits for neural systems and their evolution, and this paper aims to improve the evolutionary neural network algorithm EPNet to take advantage of those benefits. Neural networks exist with varying degrees of modularity ranging from pure modular networks characterized by disjoint partitions of hidden nodes with no communication between modules, to pure homogeneous networks with significant connections throughout. In between are apparently homogeneous networks that can be seen to have some degree of modularity if the hidden nodes are reorganized appropriately. In this paper, a modularity measure is presented and extended that can be applied to any neuron at any level in the network to provide a fine analysis of node partitioning. It also allows the rearrangement of nodes to create modules in homogeneous networks, and that is used to improve the EPNet algorithm to evolve modular neural networks. Experimental results on a simple classification task confirm that the new modular EPNet algorithm does indeed lead to more modular networks than the classical EPNet algorithm, without compromising the performance on the given task.

This work provides an analysis of using the evolutionary algorithm EPNet to create ensembles of artificial neural networks to solve a range of forecasting tasks. Several previous studies have tested the EPNet algorithm in the classification field, taking the best individuals to solve the problem and creating ensembles to improve the performance. But no studies have analyzed the behavior of the algorithm in detail for time series forecasting, nor used ensembles to try to improve the predictions. Thus, the aim of this work is to compare the ensemble approach, using two linear combination methods to calculate the output, against the best individual found. Since there are several parameters to adjust, experiments are set up to optimize them and improve the performance of the algorithm. The algorithm is tested on 21 time series of different behaviors. The experimental results show that, for time series forecasting, it is possible to improve the performance by using the ensemble method rather than using the best individual. This demonstrates that the information contained in the EPNet population is better than the information carried by any one individual.

This paper explores the idea that auto-teaching neural networks with evolved selfsupervision signals can lead to improved performance in dynamic environments where there is insufficient training data available within an individual’s lifetime. Results are presented from a series of artificial life experiments which investigate whether, when, and how this approach can lead to performance enhancements, in a simple problem domain that captures season dependent foraging behaviour.

After a review of the purpose and biological background of Head Direction Cells, and existing models of them, we introduce an improved neural model that integrates visual flow information with the vestibular inputs found in earlier models. Simulation results using real video inputs are presented to explore and validate the new model, the evolutionary advantages for the proposed arrangement are noted, its utility for use in robot controllers is discussed, and suggestions are made for further work in this area.

A series of evolutionary neural network simulations are presented which explore the hypothesis that learning factors can result in the evolution of long periods of parental protection and late onset of maturity. By evolving populations of neural networks to learn quickly to perform well on simple classification tasks, it is shown that better learned performance is obtained if protection from competition is provided during the network's early learning period. Moreover, if the length of the protection period is allowed to evolve, it does result in the emergence of relatively long protection periods, even if there are other costs involved, such as individuals not being allowed to reproduce during their protection phase, and the parents suffering increased risk of dying while protecting their offspring.

The idea of using evolutionary techniques to optimize the performance of neural networks is now widely used, but some approaches have been found to result in the evolution of risky learning strategies that lead to occasional instances of very poor performance. In this paper I shall present a series of simulations that explore the nature of those problems, and examine the extent to which one can use ensemble techniques to alleviate them.

It is well known that incremental learning can often be difficult for traditional neural network systems, due to newly learned information interfering with previously learned information. In this paper, we present simulation results which demonstrate how evolutionary computation techniques can be used to generate neural network incremental learners that exhibit improved performance over existing systems.

This paper presents the results of simulations of a new class of artificial neural network models of reading. Unlike previous models, they are not restricted to mono-syllabic words, require no complicated input-output representations such as Wickelfeatures and, although based on the NETtalk system of Sejnowski and Rosenberg (1987), require no pre-processing to align the letters and phonemes in the training data. The best cases are able to achieve 100% performance on the Seidenberg and McClelland (1989) training corpus, in excess of 90% on pronounceable nonwords and on damage exhibit symptoms similar to acquired surface dyslexia.

Attempts in the past to use evolutionary simulations to model the emergence of modular neural architectures has led to conflicting results. Here, we present some preliminary work on the use of computational embryogeny to model the evolution of neural architecture at a less coarse level of description. We believe that such an approach will lead to much more reliable and biologically realistic simulations of brain evolution.

Catastrophic forgetting is a well-known failing of many neural network systems whereby training on new patterns causes them to forget previously learned patterns. Humans have evolved mechanisms to minimize this problem, and in this paper we present our preliminary attempts to use simulated evolution to generate neural networks that suffer significantly less from catastrophic forgetting than traditionally formulated networks.

For autonomous neural network systems one usually needs fast learning and good generalization performance, and there is inevitably a trade-off between these two requirements. Using evolutionary techniques can generate high performance networks, but this often leads to unwanted side effects, such as occasional instances of very poor performance. This paper explores the problems that arise for traditional evolved neural networks using a range of evolutionary approaches, and shows how they can, to a large extent, be overcome by allowing the networks to evolve age dependent plasticities.

The use of simulated evolution is now a commonplace technique for optimizing the learning abilities of neural network systems. Neural network details such as architecture, initial weight distributions, gradient descent learning rates, and regularization parameters, have all been successfully evolved to result in improved performance. In this paper I investigate which evolutionary approaches work best in this field. In particular, I compare the traditional generational approach with a more biologically realistic steady-state approach.

A neural network model is presented which is an abstraction of many real world adaptable control systems that seems to be sufficiently complex to provide interesting results, yet simple enough that computer simulations of its evolution can be carried out in weeks rather than years. Some preliminary explorations of the interaction between learning and evolution in this system are described, together with some suggestions for future research in this area.

Several recent papers have described how lexical properties of words can be captured by simple measurements of which other words tend to occur close to them. At a practical level, word co–occurrence statistics are used to generate high dimensional vector space representations and appropriate distance metrics are defined on those spaces. The resulting co–occurrence vectors have been used to account for phenomena ranging from semantic priming to vocabulary acquisition. We have developed a simple and highly efficient system for computing useful word co–occurrence statistics, along with a number of criteria for optimizing and validating the resulting representations. Other workers have advocated various methods for reducing the number of dimensions in the co–occurrence vectors. LundB LandauerD and Lowe&McDonald [8] have used a statistical reliability criterion. We have used a simpler framework that orders and truncates the dimensions according to their word frequency. Here we compare how the different methods perform for two evaluation criteria and briefly discuss the consequences of the different methodologies for work within cognitive or neural computation.

Although much has been written on this subject, there still seems to be considerable confusion in the literature concerning dissociations, double dissociations and what they really mean, especially when connectionist or neural network models are involved. In this paper I attempt to clarify matters by looking at the subject from the point of view of patterns of learning rates in neural network models.

In order to maintain clear, single binocular vision developmentally and in different visual environments, it is necessary for the oculomotor system to be plastic. During development, plasticity is required to compensate for changes in the size of the eyeball and the distance between the eyes. Adults also show plasticity when wearing optical devices (such as spectacle lenses and head mounted displays) which require local adaptation of the oculomotor system. It is therefore of interest to study the details of this plasticity since it will have consequences for the types of optical devices that can be used with comfort and safety. We have begun this study by simulating the human adaptive response to vertical disparities at different points in the visual field, and using comparisons with empirical data to explore the possible biological limitations of this adaptation process.

Many connectionist language processing models have now reached a level of detail at which more realistic representations of semantics are required. In this paper we discuss the extraction of semantic representations from the word co-occurrence statistics of large text corpora and present a preliminary investigation into the validation and optimisation of such representations. We find that there is significantly more variation across the extraction procedures and evaluation criteria than is commonly assumed.

Connectionist models of the mapping from orthography or phonology to random binary semantic vectors allow the simulation of lexical decision with reaction times that show patterns of semantic and associative priming similar to those found experimentally with human subjects. The co-occurrence statistics of words in large corpora allow the generation of vectors whose distribution correlates with the perceived semantic relatedness of the words. Here we discuss the use of these more realistic corpus derived semantic representations in connectionist models of lexical decision. We find lexical decision priming that correlates with distances in the semantic vector space, but the reaction times are very noisy. Averages over many words and/or many networks are required for the relationships to become clear. The question of associative priming remains open.

The connectionist modeling of reading, spelling, and past tense acquisition is discussed. We show how the same simple pattern association network for all three tasks can achieve perfect performance on training data containing many irregular words, provide near human level generalization performance, and exhibit some realistic developmental and brain damage effects. It is also shown how reaction times (such as naming latencies) can be extracted from these networks along with independent priming and speed-accuracy trade-off effects. We argue that all the remaining problems with these models will be solved by supplementing them with an appropriate connectionist semantic route.

We discuss the simulation of reaction times in connectionist systems. The obvious way to do this involves thinking in terms of neural activations building up towards some threshold in cascaded systems, but it has also been suggested that the output activation error scores in standard back-propagation networks should also be correlated with response times. The idea is that in the more realistic cascaded processing systems, the clearer the outputs (i.e. the lower the errors), the lower the time taken to reach the threshold. If this is correct and we can consider our simple feedforward networks to be reasonable approximations to these cascaded systems, then we have an easy way to simulate reaction times. However, the validity of this has been questioned. I will discuss these issues in some detail, suggest a more principled way of extracting simulated reaction times from back-propagation networks and show how this relates to the corresponding cascaded networks.