Considering the lack of development for faster and more efficient solutions to the Grid World Q-Learning problems, it is no surprise that the performance of Q-Learning for Grid Worlds tends to exponentially reduce in effectiveness, as board size increases. This is a problem, because Q-Learning could (conceivably) solve analytic and mathematical problems, if runtime performance were enhanced. The motivation for this research, therefore, is to increase the runtime performance of Q-Learning for Grid Worlds by development of Multiple Awareness No-Retracing Agents for Q-Learning (or MANAQL). An account of conventional Q-Learning for Grid World environments is given and includes the anatomy of a Q-Learning class, its functions and their intended interactions; A detailed account of the methodology used to create MANAQL and how its functions and components differ from those of conventional Q-Learning is also given. In the Literature Review, several other Q-Learning algorithms that focus on multiple agents are narratively assessed and the question of why so little work has been done on improving Q-Learning for Grid World is asked and answered. The time complexity of the centred and non-centred conventional Q-Learning algorithms are evaluated and compared with each other and with those of the centred and non-centred MANAQL algorithms. Applications of the new algorithm will be the focus of later work.
This paper focuses on developing an innovative method that incorporates facial features, such as mouth, and eyebrows to identify emotions from an image. The proposed system can also recognize emotions from images with the facial pose variation and occlusions. Our findings indicate that this new system effectively identifies critical and comprehensive features. These features are then processed through a second phase involving a multi-objective optimization technique. This technique accurately predicts emotions in an image, showing superior performance compared to many standard and deep learning-based methods. Our approach is particularly adept at changes in facial angles, outperforming many conventional and advanced models. Our model's better performance in emotion recognition is due to its ability to choose the optimal solution from a range of possible solutions, which allows it to accurately represent the most suitable emotional expression seen in the face images.
Deep learning networks effectively address the challenge of transforming low-resolution images into high-resolution images by learning from a series of LR-HR sample pairs. However, most network models are specifically trained for certain scales, and each set of network parameters is only applicable to a particular scale of super-resolution problems. To address this issue, this study introduces an arbitrary-scale super-resolution neural operator network based on a Galerkin attention mechanism, integrating Residual Channel Attention Networks as a replacement for the original feature extraction module. Furthermore, it investigates the impact of different loss functions, training epochs, and feature extraction modules on the performance of the super-resolution neural operator. Experimental results validate the performance of the proposed feature extraction module. The findings indicate that, under the same loss functions and training epochs, the improved module exhibits smaller losses on the training set compared to the original module, demonstrating enhancements. Even with significantly more training epochs, the visual effects of the original network using EDSR-Baseline as the feature extraction module still fall short of those achieved by the improved network.
Transforming the energy supply to a sustainable system requires the integration of numerous renewable generation. This in turn leads to a need for new algorithms to keep the increasing complexity manageable. Support vector decoders for the systematic generation of feasible schedules for the operational management of electrical generators and consumers have proven to be helpful here. Such a decoder models a system by using a hypersphere model in a high-dimensional Hilbert space. However, it has been known for some time that the volume of high-dimensional hyperspheres exhibits an anomaly regarding the development of the volume with growing dimensionality. This paper investigates the influence of this anomaly on the performance of models based on 1-class support vector descriptions that use a high-dimensional hypersphere for modeling. We show that there is an impact that affects a certain range of support vector models, but also allows for Pareto-optimal models when training for quick execution is an issue.
While the non-linearity and sensitivity of chaotic systems make the forecasting of their future behaviour challenging, the ability of machine learning techniques to accurately represent complex dynamics has placed them at the forefront of this field. However, small differences in the model used can lead to noticeable differences in observable model performance due to the systems' sensitivity. This can result in even seemingly inconsequential changes to the training process causing practical differences in model performance. In this study performance metrics for both next-step prediction accuracy and a novel metric for the duration of accurate prediction have been calculated for a number of long-short term memory (LSTM) models, and their variance studied. The results highlight the causes and quantifies the scale of inconsistency in model performance that can be found during seemingly equivalent training scenarios, and concludes that such variance in model performance is non-trivial.
Recently, the quantum computer's ability to perform reliable computations beyond the capabilities of classical computing methods is referred to as quantum utility. To achieve this ability, it is becoming increasingly important to assess quantum algorithm performance in practical applications. Starting from this consideration, this work investigates the performance of the well-known Quantum Approximate Optimisation Algorithm (QAOA) with a Genetic Algorithm-based training in solving a real-world problem in the domain of power systems. As shown in the reported experiments using both an ideal simulator and a real IBM quantum processor, QAOA empowered by genetic algorithms outperforms the compared algorithms, particularly on real quantum hardware at the highest QAOA circuit depth.
Finding Dairy Cattle (DC) which can produce high levels of milk while emitting low levels of methane (CH4) is a key goal for agriculture. We applied two classification systems in the prediction of DC production and emission levels combined. A Multilabel system (MLS), which utilised an individual model for the prediction of each individual phenotype of the combination, and a Multiclass system (MCS), which applied a single model in the direct prediction of the phenotypes pre-combined. The mean difference between the MLS and MCS systems was not statistically significant (p > 0.05), scoring an overall average accuracy of 66
Instance space analysis extends the algorithm selection framework by enabling the visualisation of problem instances via dimensionality reduction (DR). The lower dimensional projection can also be used as input to predict algorithm performance, or to perform algorithm selection. In this paper we consider two supervised DR methods - partial least squares (PLS) and linear discriminant analysis (LDA) - both as visualisation tools and for the purpose of constructing classification models for algorithm selection. Multinomial logistic regression models are used for the classification problem. We compare PLS and LDA to DR methods previously used in this context on three combinatorial optimisation problems, and show that these methods are as competitive.
Few-shot object detection (FSOD) seeks to detect novel classes with limited training instances and has received a great deal of interest. Existing meta-learning based methods extract prototypes to aggregate with query features for instance classification and localization. However, they often suffer from lack of effective prototypes. In this paper, we propose a FSOD model, termed enhanced prototype net (EPNet), based on meta-learning. We first present an early stage aggregation module, which aggregates cross-attention query features and support features to enhance prototypes generating. Then a feature convex combination mechanism is proposed to perform data augmentation in the feature space and a soft distance loss to shrink the regions of prototypes features, thus extracting distinct and robust prototypes. Besides, a class-agnostic aggregation module is developed to fuse the query features with every prototype to boost training query samples for instance classification. Experimental results on Pascal VOC and MS COCO datasets demonstrate that the proposed model surpassed the baseline model and achieved state-of-the-art performances.
As the elderly population grows, fall prediction and prevention becomes a crucial subject for investigation. This work explores a novel approach using unobtrusive sensor such as pressure mats and machine learning (ML) algorithms to continuously monitor gait patterns and predict fall risks in older adults. Sensing pressure mat was denoised and used to acquire movement datasets from an individual ranging from normal walking to performance of various activities such as falls, jumps and the application of uneven foot pressure while walking or exercising in a lab environment that mimics a home environment. Data gleaned from the sensor were cleaned, visualised, analysed and used to design a fall prediction model using a decision tree algorithm. Experimental results indicated a balanced performance (overall accuracy of 80
In academic institutions, among the most important success criteria is students’ academic performance. However, one of the biggest challenges facing institutions has been the early detection and enhancement of students’ academic performance at all levels. Students may run into several issues that hinder their ability to study, thereby, having a detrimental effect on their academic achievement. These problems can be effectively resolved if student data is pre-analyzed with early performance predictions, to enable prompt support decisions. Thus, this work applied an Adaptive Neuro Fuzzy Inference System (ANFIS) with subtractive clustering to predict students’ academic performance and identify factors that influences the students’ performance. Hence, this would be helpful in making informed decisions that support students who require assistance; also, taking effective steps to improving their academic performance. Furthermore, due to the benefits of mixing both neural networks and fuzzy systems, the applied ANFIS model, which is a hybrid learning algorithm, processes information quickly to produce more comprehensible and interpretable insight. Also, subtractive clustering (SC) was used to group similar characteristics in the dataset, to decrease the number of rules and membership functions of ANFIS, thereby, reducing the complexity of ANFIS. Academic records of computer science students at the University of Huddersfield from 2017–2022 were used, which provided several features useful in predicting the students’ performance. Evaluating the results of ANFIS-SC with recommended machine learning techniques in articles showed that Decision Tree, Ada Boost Regression, and Neural Networks have a high accuracy score like the Adaptive Neuro Fuzzy Inference System with Subtractive Clustering (ANFIS-SC).
The emergence of neural architecture search (NAS) technology has lowered the professional threshold for optimizing model architectures. However, existing NAS methods primarily evaluate performance by fully training a network architecture, which is computationally expensive and slow. This paper proposes a hybrid performance estimation strategy search framework for neural architecture search, which can flexibly adjust the performance evaluation strategy at each stage. In the initial stage, this study uses less accurate but low-cost methods to quickly eliminate suboptimal architectures. As the search progresses, more computationally intensive but accurate evaluation strategies are employed to filter out the optimal network architectures. In the final stage, more precise verification is conducted to ensure that the selected network architecture achieves the best performance in practice. This research can adapt to different precision and speed requirements, providing flexible reduction space ratio strategies aimed at meeting accuracy requirements while maintaining efficiency. Its generalizability and flexibility help address various NAS challenges. Experimental results show that the method proposed in this study performs excellently in multiple benchmark tests, achieving a balance between performance and efficiency. Additionally, by testing on other search spaces, datasets, and tasks, this study demonstrates its good generalization ability.
This paper proposes a new learning algorithm, Variational Generative Predictive Coding (VGPC), for generating new samples from a given dataset. VGPC learns using locally computed errors and uses in-parallel information propagation across the network. VGPC is based on Predictive Coding (PC) which is a computational theory for processing information in the brain. PC treats the brain as a generative model. The key concept of PC is that the neuronal activities in a given layer predict the neuronal activities in the next layer. The errors between the current neuronal activities of a layer and its corresponding predictions are used to infer better representations for a layer and the weights. The approach does not require a systematic feedforward and feedback propagation for prediction and learning as needed by methods trained using error-backpropagation. The representations and weights associated with any given layer are updated parallelly across the network using locally computed errors. We extend PC with an architecture for variational inference and incorporate the KL-divergence loss for learning in the bottleneck layer. The generation performance of VGPC, evaluated with MNIST, FashionMNIST and CelebA datasets, is compared with generative methods based on Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) using two metrics, called Fréchet Inception Distance (FID) and Inception Score (IS). The results indicate that VGPC achieves higher generation performance compared to GAN and VAE-based methods on MNIST and FashionMNIST datasets, with FID score reductions ranging from 47.0
In this paper we propose a light-weight disperse approach to detection of cyber threats at the network edge. With each edge node acting in isolation and independent of other transit nodes, with a limited perspective of the entire traffic targetting a destination host. Using a combination of three feature selection algorithms (hence reducing models computational overhead) as well as nine machine learning and deep learning classifiers, this research investigates suitable combinations that meet the following criteria: 1. suitably light-weight (requiring minimal compute resources), 2. having suitably high DDoS detection performance, 3. suitable for isolated source-end detection, and 4. able to detect DDoS attacks that utilise traffic flows which are indistinguishable from user traffic. Detailed performance comparisons of centralised detection is carried out against edge-detection by using a fraction of the training dataset available to the centralised model, hence simulating the view of edge nodes. To investigate detection performance of classifiers when DDoS attacks uses traffic similar to regular network data, a virtual network was implemented and utilised to generate two unique datasets. This paper also details the impact of using stateless benign flows for DDoS attacks compared to using stateful flows. The results clearly indicate the feasibility of our framework using light-weight disperse models to detect DDoS traffic in cases where known malicious and stateless benign network packets are used in the DDoS attack.
Commercial waste collection can be modelled as a vehicle routing problem with a high number of stops per route, corresponding to bins from individual customers. Retail collections may occur in pedestrian precincts, where access is restricted by time of day. Many commercial collections, particularly from retail areas, occur in highly congested zones, such as high streets. Therefore, modelling with time-of-day dependent travel speeds and turning time penalties (e.g., turning right onto a main road) is essential for accurate time estimation. This study aims to investigate heuristics to solve this problem, specifically using a cluster-first, route-second approach for construction heuristics based on graph partitioning of the road network. Problem instances have been generated, and promising results have been achieved.
Training deep learning models for medical applications often requires manual annotation of data at scale, which can become impossible under circumstances as annotated medical data is often scarce and limited. Yet, availability of abundant unlabelled data can be exploited for the development of better deep models. The proposed approach leverages a bipartite strategy, where a contrastive pre-training method is employed to enhance the representation of the backbone network, followed by fine-tuning with limited labeled data. Experiments on several medical detection tasks were conducted. Results showed significant performance improvements, with our method achieving 45.67
Accurate segmentation of teeth and gums plays a pivotal role in dental imaging, influencing both diagnostic accuracy and the efficacy of subsequent treatment strategies. Previous research predominantly focused on single feature methodologies which limits their ability to capture the complex and nuanced structures present in dental images. This study proposes a feature fusion approach to enhance segmentation performance by integrating multiple features and leveraging their complementary strengths. We introduce a multi-feature model that integrates a diverse set of features, encompassing intrinsic properties such as curvature and density, alongside texture information from Spin images, and local shape descriptors from Signature of Histograms of Orientations (SHOT) and Fast Point Feature Histogram (FPFH). We have developed a machine learning model that adeptly captures the intricate geometrical nuances of the oral cavity. The fusion of these features enables a dual focus on both the broader shape and the finer details, ensuring a thorough representation of dental structures. Our experiments showed that the feature fusion approach significantly improves segmentation accuracy and robustness. This comprehensive evaluation, which encompasses feature ablation studies and rigorous cross-validation, validates the superior performance of the model compared to traditional single feature methodologies. The optimal feature combination of Curvature, Density, and FPFH descriptors achieved an accuracy of 94.19
Change Detection in remote sensing images typically aims to accurately determine any significant land surface changes, based on acquired multi-temporal image data, being a pivotal task in remote sensing image processing. Recently, deep learning has been widely applied in machine vision with remarkable potential demonstrated for performing change detection in images. Current multi-scale feature fusion methods, while enhancing algorithm performance, often introduce a significant number of redundant parameters, thereby increasing model complexity. This paper presents a novel approach to addressing this challenge. A cross-scale heterogeneous convolution change detection method is proposed. It strengthens the perception of multi-scale information without adding excessive redundant parameters, mitigating the detail-blurring issues caused by fusion. By integrating scale perception with spatial-spectral information aggregation, the proposed approach effectively alleviates scale sensitivity issues, improving change detection performance in complex multi-scale environments. This is showcased by comparative experiments on a challenging real-world dataset with seven existing high-performing methods, the proposed method achieved an F1-score of 84.15
Automation is at the core of current technology development, and it is becoming increasingly prevalent, hence the need to investigate and develop simple systems/algorithms for dependable and efficient implementations. Autonomous robots designed to navigate unfamiliar environments successfully are often needed in automation systems. There have been many approaches to autonomous robot navigation systems. These include single-robot navigation and multi-robot coordination systems. Simplicity, effectiveness, and efficiency are paramount factors to be considered as this field grows. Multi-robot coordination is now a popular subject in robotism. This research conducted a constructive investigation of bug algorithms, and then developed and implemented a technique for optimizing some of the current algorithms. Two advanced simple navigation algorithms, with high effectiveness and efficiency, were thereby developed by this project. One is a single-robot algorithm which is an improvement to an existing algorithm. The other is a dual-robot navigation system (based on two different algorithms) that could be implemented for any number of robots.
Reservoir computing is a framework which uses the non-linear internal dynamics of a recurrent neural network to perform complex non-linear transformations of the input. This enables reservoirs to carry out a variety of tasks involving the processing of time-dependent or sequential-based signals. Reservoirs are particularly suited for tasks that require memory or the handling of temporal sequences, common in areas such as speech recognition, time series prediction, and signal processing. Learning is restricted to the output layer and can be thought of as "reading out" or "selecting from" the states of the reservoir. With all but the output weights fixed they do not have the costly and difficult training associated with deep neural networks. However, while the reservoir computing framework shows a lot of promise in terms of efficiency and capability, it can be unreliable. Existing studies show that small changes in hyperparameters can markedly affect the network's performance. Here we studied the role of network topologies in reservoir computing in the carrying out of three conceptually different tasks: working memory, perceptual decision making, and chaotic time-series prediction. We implemented three different network topologies (ring, lattice, and random) and tested reservoir network performances on the tasks. We then used algebraic topological tools of directed simplicial cliques to study deeper connections between network topology and function, making comparisons across performance and linking with existing reservoir research.
