We present a topological framework for analyzing neural time series that integrates Transfer Entropy (TE) with directed Persistent Homology (PH) to characterize information flow in spiking neural systems. TE quantifies directional influence between neurons, producing weighted, directed graphs that reflect dynamic interactions. These graphs are then analyzed using PH, enabling assessment of topological complexity across multiple structural scales and dimensions. We apply this TE+PH pipeline to synthetic spiking networks trained on logic gate tasks, image-classification networks exposed to structured and perturbed inputs, and mouse cortical recordings annotated with behavioral events. Across all settings, the resulting topological signatures reveal distinctions in task complexity, stimulus structure, and behavioral regime. Higher-dimensional features become more prominent in complex or noisy conditions, reflecting interaction patterns that extend beyond pairwise connectivity. Our findings offer a principled approach to mapping directed information flow onto global organizational patterns in both artificial and biological neural systems. The framework is generalizable and interpretable, making it well suited for neural systems with time-resolved and binary spiking data.

Abstract Rail freight planning problems pose specific challenges that have attracted the attention of academics and industry professionals for many decades. They involve multiple types of assets (trains, stations, terminals, etc.) and are subjected to structural, operational and safety constraints. Even though various approaches have been proposed, few can address the complexity and size of real‐world scenarios, and decentralized techniques, like multi‐agent systems (MAS), have become more prevalent. The current state of the art in disciplines such as agent technology, reinforcement learning and discrete‐event simulation allows the implementation of complex architectures, with multiple actors interacting and learning simultaneously. Therefore, this study takes advantage of these current advances and proposes an innovative approach to real‐time traffic management problems in freight railway networks through multi‐agent deep reinforcement learning (MADRL). This study was motivated by the decision‐making scheduling problems arising in the Hunter Valley Coal Chain (HVCC), located in New South Wales, Australia. The MADRL algorithm uses as the training environment the simulation model currently utilized for capacity planning of the HVCC, allowing experiments with actual data. Thus, we enhanced the simulation model to accommodate a MAS with intelligent agents representing system elements, such as trains, dump stations, and load points. Furthermore, these agents act in a decentralized fashion based on local observations, constituting a partially‐observed Markov decision process (dec‐POMDP). Three variations of the MADRL approach are presented: a baseline model, an extended model, and one that directly addresses deadlocks. Finally, we present a transfer learning method that improves deadlock resolution and leverages performance. In the experiments, we explore specific, complex scenarios arising in the HVCC, where trains frequently face deadlock conditions. The baseline model outperforms a first‐come‐first‐serve (FCFS) based heuristic used by HVCC's simulation model and a genetic algorithm in instances with up to 60 trains – but fails in more complex scenarios. On the other hand, the most advanced model, which addresses deadlocks via transfer learning, always finds feasible solutions and produces policies that outperform the FCFS‐based heuristic in 94% of the instances.

The topological analysis of four-dimensional (4D) image-type data is challenged by the immense size that these datasets can reach. This can render the direct application of methods, like persistent homology and convolutional neural networks (CNNs), impractical due to computational constraints. This study aims to estimate the topology type of 4D image-type data cubes that exhibit topological intricateness and size above our current processing capacity. The experiments using synthesised 4D data and a real-world 3D data set demonstrate that it is possible to circumvent computational complexity issues by applying downscaling methods to the data before training a CNN. This is achievable even when persistent homology software indicates that downscaling can significantly alter the homology of the training data. When provided with downscaled test data, the CNN can still estimate the Betti numbers of the original sample cubes with over 80% accuracy, which outperforms the persistent homology approach, whose accuracy deteriorates under the same conditions. The accuracy of the CNNs can be further increased by moving from a mathematically-guided approach to a more vision-based approach where cavity types replace the Betti numbers as training targets.

This paper proposes a low-complexity data-driven multi-user detector for grant-free non-orthogonal multiple access (GF-NOMA), which has gained significant interest in Internet of Things (IoT). IoT traffic is predominantly sporadic, where devices become active whenever they have data to transmit. The conventional grant-access procedure for requesting a transmission slot every time results in significant signaling overhead and latency. In power domain GF-NOMA, multiple devices can be preallocated the same channel resource, but different power levels. Whenever a device has data, it starts transmission directly using the allocated power level without any grant request. While this significantly reduces the signaling overhead, the access point has to perform the complex task of identifying the active devices and decoding their data. Conventional receivers for power domain NOMA fail in such GF scenarios and the typical solution is to limit transmissions to be packet-synchronized and add carefully chosen pilots in every packet to facilitate activity detection. However, in fairly static IoT networks with low-complexity devices and small packet sizes, this represents a significant overhead and reduces efficiency. In this work we solve the GF-NOMA detection problem without these constraints, by analyzing the boundaries of the received constellation points in power domain GF-NOMA for all activation combinations at once. A low-complexity decision tree-based receiver is proposed, which performs as well as the maximum likelihood-based benchmark receiver, and better than traditional data-driven detectors for GF-NOMA. Comprehensive simulation results demonstrate the performance of the proposed detector in terms of its detection efficiency and parameter learning with minimal training data.

Zero-shot learning is a scenario in machine learning where the classes used in the training and test sets are disjoint. This work considers zero-shot learning for environmental audio and improves results by enhancing audio and word embeddings. Previous works use the VGGish model for audio embeddings, and textual class labels are often used as input for word embedding networks such as Word2Vec. This study instead uses a modified YAMNet network to obtain semantic audio embeddings for zero-shot learning. Moreover, part of this study involves adding linguistic devices, such as synonyms, semantic broadening and onomatopoeia, to the input of the word embeddings. With these two modifications, top-1 accuracy is increased on average by over five percentage points compared to the state-of-the-art on ESC-50. This emerging area of research has applications in robot awareness, security systems and wildlife conservation in situations where no data is available for some classes.

Deep Neural Networks (DNN) require specifically tuned architectures and hyperparameters when being applied to any given task. Nature-inspired algorithms have been successfully applied for optimising various hyperparameters in different types of DNNs such as convolutional and recurrent for sentence classification. Hybrid networks, which contain multiple types of neural architectures have more recently been used for sentence classification in order to achieve better performance. However, the inclusion of hybrid architectures creates numerous possibilities of designing the network and those sub-networks also need fine-tuning. At present these hybrid networks are designed manually and various organisation attempts are noticed. In order to understand the benefit and the best design principle of such hybrid DNNs for sentence classification, in this work we used an Evolutionary Algorithm (EA) to optimise the topology and various hyperparameters in different types of layers within the network. In our experiments, the proposed EA designed the hybrid networks by using a single dataset and evaluated the evolved networks on multiple other datasets to validate their generalisation capability. We compared the EA-designed hybrid networks with human-designed hybrid networks in addition to other EA-optimised and expert-designed non-hybrid architectures.

The performance that Deep Neural Networks can achieve on a specific task is impacted significantly by the hyperparameters selected. In order to compare the performance of various Deep Neural Network architectures on sentence classification, an optimised set of hyperparameters needs to be found for each architecture. In this work we use a simple Genetic Algorithm to optimise the hyperparameters of three different architectures and we evaluate their performance on a suite of sentence classification benchmarks. We found that a single Genetic Algorithm is capable of optimising a variety of different architectures and that the evolved configurations found can compete with those chosen by experts while using fewer overall trainable parameters. The three architectures tested are a recurrent neural network and two types of convolutional networks with a large difference in complexity. Of the three architectures, optimised for sentence classification, a simple Convolutional Neural Network was the overall best performer consistently achieving good performance while using very few trainable parameters.

In this paper, we propose to use deep reinforcement learning (DRL) for the task of cooperative spectrum sensing (CSS) in a cognitive radio network. We selected a recently proposed offline DRL method called conservative Q-learning (CQL) due to its ability to learn complex data distributions efficiently. The task of CSS is performed as follows. Each secondary user (SU) performs local sensing and using CQL algorithm, determines the presence of licensed user for current and$k$-1 future timeslots. These results are forwarded to the fusion centre where another CQL algorithm is operating that generates a global decision for the current and$k$-1 future timeslots. Then, SUs do not perform sensing for the next$k$-1 timeslots to save energy. The proposed CSS mechanism can significantly increase the licensed user detection accuracy and the data transmissions by SUs. In addition, it reduces the sensing results transmission overhead. The proposed solution is tested with a stochastic traffic load model for different activity patterns. Our simulation results show that the proposed problem formulation using the CQL algorithm can achieve similar detection accuracy as other state-of-the-art methods for CSS while significantly reducing the computation time.

Robust optimization of real-world problems is essential to reduce the significant negative impact of uncertainties and noises present in the environment. Uncertainties in the decision variables are often handled using explicit or implicit averaging methods, in which the fitness of a solution isjudged based on the objective values of neighbouring solutions. Explicit averaging methods are highly reliable but require additional objective function evaluation, which can significantly increases the overall computational cost of an optimization process. On the other hand, implicit averaging techniques are computationally cheap, yet they suffer from low reliability since they use the history of search in a population-based optimization algorithm. This work proposes a conditional Pareto optimal dominance to improve the reliability of robust optimization methods that use implicit averaging methods. The proposed method is applied to Multi-Objective Particle Swarm optimisation. Empirical study with a benchmark suite shows the benefit of the proposed conditional Pareto optimal dominance in locating robust solutions in multi-objective problems.

We propose a speech emotion recognition system based on deep neural networks, operating on raw speech data in an end-to-end manner to predict continuous emotions in arousal-valence space. The model is trained using time and frequency information of speech recordings of the publicly available part of the multi-modal RECOLA database. We use the Concordance Correlation Coefficient (CCC) as it was proposed by the Audio-Visual Emotion Challenges to measure the similarity between the network prediction and gold-standard. The CCC prediction results of our model outperform the results achieved by other state-of-the-art end-to-end models. The innovative aspect of our study is an end-to-end approach to using data that previously was mostly used by approaches involving combinations of pre-processing or post-processing. Our study used only a small subset of the RECOLA dataset and obtained better results than previous studies that used the full dataset.

In this paper, we investigate the task of quality of service (QoS) routing in software defined networks (SDN). We consider delay, bandwidth, loss, and cost as QoS parameters. We propose a new deep reinforcement learning solution for greedy online QoS routing in SDN and call it Deep Q-Routing (DQR). DQR utilises a dueling deep Q-network with prioritised experience replay to compute a path for any source-destination pair request in the presence of multiple QoS metrics. In contrast to existing DRL-based routing methods, the proposed DQR method regards the task of routing as a discrete control problem and uses a reward function comprising weighted QoS parameters. Our simulation results show that DQR substantially improves end-to-end throughput compared to other existing learning based methods.

This paper investigates the task how to achieve fairness in distributed dynamic spectrum access (DSA). Specifically, we consider a cognitive radio network scenario with multiple primary users (PUs) and secondary users (SUs). Each PU operates in a licensed channel. We assume that there is no coordination between PUs and SUs, and no coordination among SUs. The key challenges for SUs are to: (1) avoid collisions with PUs, (2) avoid collisions with other SUs, (3) fair access of spectrum resources in an uncoordinated system, (4) deal with different PU activity patterns, (5) deal with spectrum sensing errors. To address these challenges, we propose a deep reinforcement learning (DRL) approach and an associated reward function to achieve fair access to spectrum resources. Specifically, we use the method of Dueling Double Deep Q-Networks with Prioritised Experience Replay (D3QN-PER) as DRL algorithm for each SU. In our simulation experiments, we demonstrate that the proposed approach performs better than existing DRL methods.

Convolutional Neural Networks (CNNs) are currently the most prominent deep neural network models and have been used with great success for image classification and other applications. The performance of CNNs depends on their architecture and hyperparameter settings. Early CNN models like LeNet and AlexNet were manually designed by experienced researchers. The empirical design and optimisation of a new CNN architecture require a lot of expertise and can be very time-consuming. In this paper, we propose a genetic algorithm that can, for a given image processing task, efficiently explore a defined space of potentially suitable CNN architectures and simultaneously optimise their hyperparameters. We named this fast automatic optimisation model fast-CNN and employed it to find competitive CNN architectures for image classification on CIFAR10. In a series of comparative simulation experiments we could demonstrate that the network designed by fast-CNN achieved nearly as good accuracy as some of the other best network models available but fast-CNN took significantly less time to evolve. The trained fast-CNN network model also generalised well to CIFAR100.

The imagination of non-existent faces in random patterns, clouds and rock formations is known as facial pareidolia. We show that facial pareidolia also occurs naturally in a standard Convolutional Neural Network (CNN) trained on face recognition. For achieving this we propose a new method to analyse CNNs that combines feature visualisation and dimensionality reduction methods to cluster the hidden neuron activations in convolutional layers into groups with discriminative roles. The main contributions of the present paper are 1.) an approach that uses a CNN trained on human face detection for facial pareidolia simulation without any additional training on a target image set of abstract facial patterns and 2.) a novel way of improving the generalisation capacity of a CNN for cross-depiction recognition and domain adaptation scenarios using features learned by hidden neurons.

While deep learning has achieved state-of-the-art results on many computer vision tasks it is still challenged when interpreting human facial expressions, namely: poor generalisation ability of models across datasets, failure to account for individual differences in similar emotional states, and inability to recognise compound facial expressions and low-intensity or subtle emotional states. This study analyses how the resolution of face images that are input to various Convolutional Neural Network (CNN) models impacts on their ability to recognise compound and low-intensity emotions. Several high-resolution facial expression databases were combined to compile a simple dataset containing high-intensity emotions and a complex data set consisting of compound and low-intensity emotions. In the experiments, standard pre-trained CNN models that were further fine-tuned achieved higher validation accuracies than CNN models that were trained from scratch on the simple data set. However, when tested on the complex data the models trained from scratch generalised better than the refined pre-trained models. Using a technique of output visualisation we could show how our high-resolution CNN models were able to generalise to the complex data where they utilised small facial features that previously were not detectable.

Manifold learning can only be successful if enough data is available. If the data is too sparse, the geometrical and topological structure of the manifold extracted from the data cannot be recognised and the manifold collapses. In this paper we used data from a simulated two-dimensional double pendulum and tested how well several manifold learning methods could extract the expected manifold, a two-dimensional torus. The experiments were repeated while the data was down sampled in several ways to test the robustness of the different manifold learning methods. We also developed a neural network-based deep autoencoder for manifold learning and demonstrated that it performed in most of our test cases similarly or better than traditional methods such as principal component analysis and isomap.

This paper proposes an efficient computational approach for estimating the topology of manifold data as it may occur in applications. For two- or three-dimensional point cloud data, the computation of Betti numbers using persistent homology tools can already be computationally very expensive. We propose an alternative approach that employs deep learning to estimate Betti numbers of manifolds approximated by point clouds. A critical aspect in this new approach is the generation of suitable synthetic training data of scalable topological complexity. Once deep neural networks are trained on this data, inference can be computationally efficient and robust to noise. The pilot results of our study for two- and three-dimensional data support the hypothesis that deep convolutional neural networks can estimate Betti numbers of simulated data that has a topological complexity beyond immediate human visual comprehension. The approach could be generalised beyond estimating the numbers of holes, cavities and tunnels in low-dimensional manifolds to counting high-dimensional holes in high-dimensional data.

Fully convolutional neural networks have shown remarkable success in performing semantic segmentation. The use of convolutional layers for the entire architecture and skip connections to combine different resolution features or predictions have been adopted in successful networks, such as U-Net and DenseNet. However, these models employ several max-pooling layers that cause the network to lose spatial information and require them to mimic an autoencoder architecture to perform semantic segmentation at the original input resolution. In this paper, we propose a network that extracts features automatically with convolutional layers, like the fully convolutional neural network, but retains the spatial information of each of the extracted features. It then utilises the extracted features to make predictions with an efficient upsampling method. We evaluate the network performance on a liver segmentation task where it performs with comparable accuracy to other state-of-the-art networks while being much smaller in terms of the number of parameters as well as faster in computation time.

RoboCup Junior is a project-oriented educational initiative that sponsors regional, national and international robotic events for young students in primary and secondary school. It leads children to the fundamentals of teamwork and complex problem solving through step-by-step logical thinking using computers and robots. The Faculty of Engineering and Built Environment at the University of Newcastle in Australia has hosted and organized the Hunter regional tournament since 2012. This paper presents an analysis of data collected from RoboCup Junior in the Hunter Region, New South Wales, Australia, for a period of six years 2012–2017 inclusive. Our study evaluates the effectiveness of the competition in terms of geographical spread, participation numbers, and gender balance. We also present a case study about current university students who have previously participated in RoboCup Junior.

Measuring the performance of an aeromagnetic compensation system is usually difficult. The standard deviation of the signal has been used as an index in the industry. While the standard deviation is drawn from frequency statistics, it cannot represent the performance on a single sampling point. On the other hand, as the true geomagnetic intensity is unknown, the signal's deviation is actually an approximate measurement of the residual error. This letter first analyzes the traditional neural model for aeromagnetic compensation to reveal the fact that the model can only estimate the expectation of interference. Then, we introduce a stochastic hidden variable to predict the standard deviation synchronously. The proposed model is derived from variational inference and trained as a stochastic gradient variational Bayes estimator. Simulations are performed to show the correlation between the true residual error and the estimated standard deviation.

This paper presents the implementation details of a proposed solution to the Emotion Recognition in the Wild 2017 Challenge, in the category of group-level emotion recognition. The objective of this sub-challenge is to classify a group's emotion as Positive, Neutral or Negative. Our proposed approach incorporates both image context and facial information extracted from an image for classification. We use Convolutional Neural Networks (CNNs) to predict facial emotions from detected faces present in an image. Predicted facial emotions are combined with scene-context information extracted by another CNN using fully connected neural network layers. Various techniques are explored by combining and training these two Deep Neural Network models in order to perform group-level emotion recognition. We evaluate our approach on the Group Affective Database 2.0 provided with the challenge. Experimental evaluations show promising performance improvements, resulting in approximately 37% improvement over the competition's baseline model on the validation dataset.

This paper proposes a simple and fast method for adapting colour lookup tables to lighting changes in real-time. The method adjusts the classified colour space regions keeping both their surface area and volume constant. Two variations of the method were compared and tested in a RoboCup soccer setting. Detection success rate was measured as a function of the speed and magnitude of hue change to the lighting environment. Compared to a static lookup table, these experimental results show improved robustness against lighting changes for detection of coloured objects.

This article presents a computer vision approach that can detect and classify abstract face-like patterns, including subliminal faces within a scene. This can be regarded as a way of simulating the phenomenon of pareidolia, that is, the tendency of humans to `see faces' in random structures such as clouds or rocks. The paper describes the system consisting of a component-based face detector and an expression classifier. The face detector creates a number of component images from the original image at different resolutions. A component image is a binary edge image where the edges are segmented into components using a labelling method with a border-following technique. The component images are then overlaid to produce a component height map where large and notable components across all resolutions have high values, while specular and noisy components have low values. The method retains three-shape components, representing two eyes and a mouth, that have height map values that are larger than the noise cut-off value. Support vector machines using scale-invariant feature vectors are applied for ranking these three-shape components by their geometry and size, and their shape semblance to human faces in the training data. The outcome is a facial expression analysis system that uses face components, with the potential to estimate an emotional expression value for a scene by producing an array of emotion scores corresponding to Ekman's seven Universal Facial Expressions of Emotion. An advantage of this technique, when compared to a holistic method, is that the face components are explicitly isolated. This supports a process of abstraction that can facilitate the detection of distorted and minimal face-like patterns.

This paper describes and experimentally evaluates a new variation of multiclass classification using support vector machines. The technique, called pairwise adaptive support vector machines (pa-SVM), is a one-vs-one multiclass classifier with each binary classifier optimized towards using the best (C,γ) parameter pair to obtain the best correct classification rate. An exponential grid search and a 10-fold cross validation algorithm was used to determine the best (C,γ) pair. To evaluate multiple (C,γ) pairs for each binary classifier with the same best correct classification rate, four scenarios of C and γ were explored (min C min γ, min C max γ, max C min γ and max C max γ). Each experiment used the radial basis function (RBF) kernel for training, and the results were obtained on 23 real world datasets from the UCI Machine Learning Repository. The results show that the pa-SVM approach mostly outperforms the standard approach and by selecting the max C min γ scenario the number of support vectors is minimized. Furthermore, the comparison of our results with recent studies using the same datasets show that the new technique is very competitive.

This paper presents an analysis system to find the impact of architectural designs on pedestrian behavioural data. The system employs GDTW-P-SVMs which are capable of modelling sequential data with variable-length input series. We apply GDTW-P-SVMs to simulated pedestrian spatial behaviour data. The data include four types of behavioural characteristics: i) movement trajectories, ii) walking speed, iii) the angle α between the movement vector and the gaze vector and iv) its derivative. The analysis system learns a statistical model characterising three classes of spatial behaviour. The classes are formed based on pedestrians' reactions to visual attractions in a simulated environment. A separate data set that includes the crowd attraction effect is used to discuss the impact of social group formation on the classification result. Our experiments show that using the angle α and its derivative as input to the classifiers results in lower classification error rates compared to classification of trajectory and speed of movement data. We compare the classification accuracy of the GDTW-P-SVMs with other classification methods that are capable of handling data objects with variable-length input series. GDTW-P-SVMs showed promising results in classifying the simulated behavioural data.

Current approaches to facial expression classification employ a variety of expression classes and different preprocessing steps, making comparison of results difficult. To outline the effects of these variations we explore several image and action preprocessing steps, using the discrete expressions: happy, sad, surprised, fearful, angry, disgusted and neutral; with a dataset aligned and normalised by our proposed face model. Each of the preprocessing steps is organised across four prominent approaches: holistic, holistic action, component and component action. These are compared using a modified multiclass Support Vector Machine (SVM) that uses pairwise adaptive model parameters. We illustrate that including the neutral expression as part of the study has a noticeable impact, and suggest that it should be used in future research in this area. We also show that results can be improved through innovative use of image and action preprocessing steps. Our best correct classification rate was 98.33% using 10-fold cross validation and a component action approach.

This paper presents the results of experiments in applying a spiking neural network to control the locomotion of a simulated biped robot. The neural model used in simulations was developed to allow for an analytic solution to a neuron's fire time, while maintaining a non-instant post-synaptic potential rise time. The synaptic weights and delays were tuned using an evolution-strategy. Simulation experiments demonstrate that already within about two thousand generations the biped is able to acquire a dynamic walk which allows it to walk upright for several metres.

Sound-scapes are useful for understanding our surrounding environments in applications such as security, source tracking or understanding human computer interaction. Accurate position or localisation information from sound-scape samples consists of many channels of high dimensional acoustic data. In this paper we demonstrate how to obtain a visual representation of sound-scapes by applying dimensionality reduction techniques to a range of artificially generated sound-scape datasets. Linear and non-linear dimensionality techniques were compared including principle component analysis (PCA), multi-dimensional scaling (MDS), locally linear embedding (LLE) and isometric feature mapping (ISOMAP). Results obtained by applying the dimensionality reduction techniques led to visual representations of affine positions of the sound source on its sound-scape manifold. These displayed clearly the order relationships of angles and intensities of the generated sound-scape samples. In a simple classification task with the artificial sound data, the successful combination of dimensionality reduction and classifier methods are demonstrated.

This study is part of a project which investigates computational principles which underlie perception and representation of architectural streetscape character. Some of the principles can be associated with fundamental concepts in brain theory and Gestalt psychology. For the experimental analysis streetscapes were represented by sequences of digital images of house facades which were prepared by a team of researchers from architecture. Two methods for non-linear dimensionality reduction, isomap and maximum variance unfolding, were applied to a set of Hough arrays (for lines) of the given images. An analysis of the extracted "streetmanifolds" revealed groupings of house facades with similar visual character and proportions. Comparative tests were conducted on a simple cylinder shaped example manifold to evaluate the geometric stability of the two dimensionality reduction methods. All experiments addressed variations of the distance metric and the neighbourhood parameter

The two-spiral task is a well-known benchmark for binary classification. The data consist of points on two intertwined spirals which cannot be linearly separated. This article reviews how this task and some of its variations have significantly inspired the development of several important methods in the history of artificial neural networks. The two-spiral task became popular for several different reasons: (1) it was regarded as extremely challenging; (2) it belonged to a suite of standard benchmark tasks; and (3) it had visual appeal and was convenient to use in pilot studies. The article also presents an example which demonstrates how small variations of the two-spiral task such as relative rotations of the two spirals can lead to qualitatively different generalisation results.

Robot learning is a growing area of research at the intersection of robotics and machine learning. The main contributions of this paper include a review of how machine learning has been used on Sony AIBO robots and at RoboCup, with a focus on the four-legged league during the years 1998-2004. The review shows that the application-oriented use of machine learning in the four-legged league was still conservative and restricted to a few well-known and easy-to-use methods such as standard decision trees, evolutionary hill climbing, and support vector machines. Method-oriented spin-off studies emerged more frequently and increasingly addressed new and advanced machine learning techniques. Further, the paper presents some details about the growing impact of machine learning in the software system developed by the authors' robot soccer team-the NUbots

Although the game of Hex has simple rules it is a challenging task for machine learning and therefore a possible testbed for incremental learning. This article first describes a possibility how to implement a simple learning artificial player for Hex. Some pilot training experiments indicate that evolutionary hill climbing is able to improve the playing strength of the player. However, the strategy to facilitate the process of learning by first using a small sized board and after some training to increase the board size seems not to work.

This article addresses the issues of colour classification and collision de- tection as they occur in the legged league robot soccer environment of RoboCup. We show how the method of one-class classification with sup- port vector machines (SVMs) can be applied to solve these tasks satisfac- torily using the limited hardware capacity of the prescribed Sony AIBO quadruped robots. The experimental evaluation shows an improvement over our previous methods of ellipse fitting for colour classification and the statistical approach used for collision detection.

Incremental learning concepts are reviewed in machine learning and neurobiology. They are identified in evolution, neurodevelopment and learning. A timeline of qualitative axon, neuron and synapse development summarizes the review on neurodevelopment. A discussion of experimental results on data incremental learning with recurrent artificial neural networks reveals that incremental learning often seems to be more efficient or powerful than standard learning but can produce unexpected side effects. A characterization of incremental learning is proposed which takes the elaborated biological and machine learning concepts into account.

This study empirically investigates variations of hill climbing algorithms for training artificial neural networks on the 5-bit parity classification task. The experiments compare the algorithms when they use different combinations of random number distributions, variations in the step size and changes of the neural networks' initial weight distribution. A hill climbing algorithm which uses inline search is proposed. In most experiments on the 5-bit parity task it performed better than simulated annealing and standard hill climbing.