Sequential perceptual aliasing is a cognitive challenge for learning agents when robots cannot differentiate states and their associations based on immediate observations, leading to poor decision-making. Existing systems struggle to abstract and distinguish observations effectively to achieve policy learning. This paper addresses this issue by introducing new aliasing types within the context of sequential aliasing and proposing an enhanced XCS classifier system that learns using a complete state-action map. The proposed system called hierarchical Frames-of-References-based XCS (Hi-FoRsXCS), can concatenate sequences of aliased states with the same observation into a chain. Hi- FoRsXCS then predicts associations between the observations and aliased states using the ends of the chain, enabling optimal policy learning with a complete action map. Experimental results demonstrate that Hi-FoRsXCS outperforms the existing systems in terms of accuracy. However, the limitations of Hi-FoRsXCS will be discussed in this paper.

The majority of computer vision algorithms fail to find higher-order (abstract) patterns in an image so are not robust against adversarial attacks, unlike human lateralized vision. Deep learning considers each input pixel in a homogeneous manner such that different parts of a ``locality-sensitive hashing table'' are often not connected, meaning higher-order patterns are not discovered. Hence these systems are not robust against noisy, irrelevant, and redundant data, resulting in the wrong prediction being made with high confidence. Conversely, vertebrate brains afford heterogeneous knowledge representation through lateralization, enabling modular learning at different levels of abstraction. This work aims to verify the effectiveness, scalability, and robustness of a lateralized approach to real-world problems that contain noisy, irrelevant, and redundant data. The experimental results of multi-class (200 classes) image classification show that the novel system effectively learns knowledge representation at multiple levels of abstraction making it more robust than other state-of-the-art techniques. Crucially, the novel lateralized system outperformed all the state-of-the-art deep learning-based systems for the classification of normal and adversarial images by 19.05% - 41.02% and 1.36% - 49.22%, respectively. Findings demonstrate the value of heterogeneous and lateralized learning for computer vision applications.

Acoustic feedback is a critical indicator for assessing the contact condition between the tool and the workpiece when humans perform grinding tasks with rotary tools. In contrast, robotic grinding systems typically rely on force sensing, with acoustic information largely ignored. This reliance on force sensors is costly and difficult to adapt to different grinding tools, whereas audio sensors (microphones) are low-cost and can be mounted on any medium that conducts grinding sound. This paper introduces a low-cost Acoustic Feedback Robotic Grinding System (AFRG) that captures audio signals with a contact microphone, estimates grinding force from the audio in real time, and enables closed-loop force control of the grinding process. Compared with conventional force-sensing approaches, AFRG achieves a 4-fold improvement in consistency across different grinding disc conditions. AFRG relies solely on a low-cost microphone, which is approximately 200-fold cheaper than conventional force sensors, as the sensing modality, providing an easily deployable, cost-effective robotic grinding solution.

Distance-based classifiers such as k-nearest neighbors remain widely used in mammographic computer-aided diagnosis and image retrieval, yet distance measure selection is rarely evaluated systematically. This paper investigates how the interaction between feature representation and distance formulation shapes mammographic lesion classification, evaluating 20 distance measures across 13 feature sets (handcrafted, deep, hybrid), three datasets (MIAS, CBIS-DDSM, INbreast), and two classifiers (kNN, nearest mean-distance), spanning 13,260 configurations under 5-fold cross-validation. Five findings emerge. First, distance choice significantly affects performance (Friedman chi(2) = 1901, p approximate to 0; best-worst AUC gap = 0.079). Second, the optimal measure depends on feature type: handcrafted features favor Learned Mahalanobis, deep features favor Adaptive Weighted Euclidean, and hybrid features favor the proposed CCFCD (AUC = 0.712). Third, measure rankings are dataset-specific (cross-dataset rho as low as -0.10). Fourth, k-NN and NMD agree at the configuration level (rho = 0.94) but not at the measure level (rho = 0.077). Fifth, increasing k improves AUC while reducing sensitivity from 0.40 to 0.14; 11.6% of configurations achieve accuracy above 80% with sensitivity below 10%, showing that AUC and accuracy alone are insufficient for clinical assessment. A computational complexity analysis shows that per-pair costs range from O(d) for most measures to O(d(2) ) for Mahalanobis, with the top-performing adaptive measures remaining tractable. These findings demonstrate that distance-based mammographic classification is governed by a coupled interaction among feature type, distance formulation, classifier, neighborhood size, and evaluation metric. Adaptive and class-aware measures consistently outperform conventional defaults when matched to the feature space.

Accuracy-based Learning Classifier Systems (LCSs) such as XCS require careful tuning of control parameters whose effective values vary by problem, limiting their applicability outside expert hands. This paper presents four knowledge-free online parameter control mechanisms for the learning rate β, error threshold ε0, subsumption threshold θsub, and don’t-care probability P#, driven solely by population-level statistics available in any accuracy-based LCS. Evaluated on XCS across 4,410 experiments (21 Boolean environments of various characteristics, including a real-world-derived design-verification problem, 30 runs each), the mechanisms reveal strongly problem-dependent effects. On overlapping problems, adaptive β and θsub break through accuracy plateaus that trap fixed-parameter XCS, lifting 11-bit Majority-On from 90.84% to 98.62%. On fully epistatic problems such as Even Parity, where all variables interact globally and no partial generalization is possible, adaptive P# is the single decisive factor, raising 5-bit Even Parity from 63.13% to 100% accuracy. However, the same mechanism causes a 37-percentage-point regression on the 20-bit Multiplexer at the standard population size, whose decomposable niche structure requires broad rules from the outset; analysis of parameter trajectories diagnoses this failure as a feedback-loop deadlock, which is fully resolved with a larger population (all runs reach 100% at N = 5000), establishing it as a population-size-dependent interaction rather than an intrinsic limitation. On DV1, a real-world-derived design-verification problem, the overlap-breaking mechanisms transfer cleanly, raising accuracy from 97.09% to 98.97% without any P#-induced regression. Notably, ε0 produces no measurable effect alone but acts as an enabler within the combined system through a feedback loop with P#. Parameter trajectories show that adaptation organizes learning into coordinated phases (early accuracy tightening, intermediate stabilization, domain-dependent generalization) rather than uniformly scaling final performance. To the best of our knowledge, this appears to be the first knowledge-free online control of P# in any LCS variant.

Emotion classification plays a crucial role in the domain of human-computer interaction, as it holds substantial significance ineffective communication. However, the task of emotion classification presents notable challenges primarily due to the subjective and multifaceted nature of emotions, encompassing varying cultural and individual interpretations. Numerous methodologies have been employed to address the complexities associated with facial emotion classification. Notably, contemporary techniques such as those that consider emotion- relevant features (e.g. mouth, eyes) have demonstrated a certain degree of success in comparison to alternative approaches (e.g. the widely employed end-to-end deep networks), which consider all pixel information in an image with equal importance. Nonetheless, a comprehensive set of guidelines evaluating not only the strengths, but also the weaknesses of deep networks remains absent. These approaches are vulnerable to even minor alterations in images, regardless of their relevance to emotion classification, e.g. a change in the foreground color of the image can reduce the accuracy of these methods by a significant amount. This results in a fragmented field, ultimately missing opportunities to enhance performance for emerging real-world applications. Therefore to address this issue, this paper offers a comprehensive review of stateof-the-art datasets and research, providing insights into different techniques and their unique contributions to emotion categorization. We also thoroughly analyze current challenges and issues, laying the groundwork for promising future research directions. Our review reveals the significance of considering contemporary methods, particularly in contrast to widely used deep networks, and their potential to enhance performance and application success.

Perceptual aliasing, where distinct states appear indistinguishable due to sensor limitations or environmental ambiguities, poses significant challenges in multi-step decision-making. The eXtended Classifier System (XCS) addresses this issue by identifying unique state transition patterns and combining them to construct accurate policies. Additionally, state-action-state chains enhance XCS's ability to handle sequentially aliased states. However, XCS processes aliased states sequentially as they are perceived, which can lead to performance degradation when incorrect versions of aliased states are included in the chain. This limitation underscores the need for a more robust mechanism to accurately differentiate unique states from aliased ones to ensure reliable policy creation. To address this, we propose a dual-stream identification framework that enhances XCS's performance in environments with perceptual aliasing. The framework introduces two parallel identification processes: one captures immediate state-action relationships, while the other identifies broader patterns across multi-step sequences. By integrating these dual streams, the proposed approach effectively disambiguates aliased states, enabling more accurate decision-making. Experimental evaluations demonstrate that our dual-stream model outperforms state-of-the-art XCS implementations across 14 benchmark environments.

Conventional classifier systems in evolutionary machine learning often struggle to detect and generalize hierarchical patterns, and thus cannot solve complex real-world problems. Inspired by lateralization and modularity in the biological brain, this extended abstract introduces the developed lateralized classifier system that evolves modular abstractions by decomposing problems into reusable knowledge components. The lateralized system consists of two complementary modules: one (Left Hemispheric Stratagem Module) focused on learning fine-grained, constituent patterns and the other (Right Hemispheric Stratagem Module) on extracting abstract, holistic relationships. Knowledge is encoded using disjunctive normal form-based code fragments and organized into hierarchical concepts stored in a heterogeneous knowledge pool. This structure enables recursive composition and facilitates efficient transfer of building-block knowledge across tasks. Experimental evaluations using complex Boolean problems demonstrate that the lateralized system significantly outperforms conventional learning classifier systems for solving complex problems such as the n-bit Parity problems and 18-bit hierarchical multiplexer problems.

Emotions are considered to convey much meaning in communication. Hence, artificial methods for emotion categorization are being developed to meet the increasing demand to introduce intelligent systems, such as robots, into shared workspaces. Deep learning algorithms have demonstrated limited competency in categorizing images from posed datasets with the main features of the face being visible. However, the use of sunglasses and face masks are common in our daily lives, especially with the outbreak of communicable diseases such as the recent coronavirus. Anecdotally, partial coverings of the face reduce the effectiveness of human communication, so would this have hampering effects on computer vision, and if so, would different emotion categories be affected equally? Here, we analyze the performance of emotion classification systems when faces are partially covered with simulated sunglasses and face masks. Deep neural networks consider all pixels in an image as equally important unlike the neuroscientific findings on how humans recognize emotions. Hence, we propose a method that considers different constituent parts (e.g. mouth, eyes, and jaw) separately, giving more attention to relevant (uncovered) regions of the face. The method is compared with three standard, partial coverings-based and attention-based methods. We found that face coverings worsen emotion categorization by up to 74% for the state-of-the-art methods, whereby emotion categories are affected differently by different coverings, e.g. clear mouth coverings have little effect on categorizing happiness, but sadness is affected badly. The proposed method (on average 60.43%) has significantly improved the performance over the standard deep learning ( $< $ 46% on average), partial coverings-based ( $< $ 47% on average), as well as the attention-based ( $< $ 51% on average) methods for both the CK+, KDEF, and RAF-DB datasets when faces were partially covered with sunglasses or different face masks.

Over the past four decades, Learning Classifier Systems (LCSs) have faced challenges in producing accurate and interpretable models for domains with continuous features, mainly due to the irrelevance issue caused by genotypic methods. These methods directly modify genotypes (conditions), leading to the creation of irrelevant rules. Phenotypic LCSs, which first modify a rule's phenotype (covered instance set) before altering its genotype, can avoid this issue. However, previous phenotypic LCSs struggle with overfitting, resulting in lower testing performance. In response, we propose a novel phenotypic LCS featuring innovations: 1) a heterogeneous phenotype approach in the rule discovery mechanism to alleviate overfitting, and 2) Informed Mutation leverages the inherent neighbouring of similar instances to enhance rule generalization, thereby improving model interpretability. The proposed LCS demonstrates its success with superior testing performance and more interpretable models in all experiments compared to other LCSs. Notably, in a problem with 2048 features, the proposed LCS model outperformed the genotypic UCS by achieving a 97.4% testing accuracy with just 13 rules, compared to the UCS's 9961 rules but only 49.9% accuracy.

Deep learning is a cutting-edge methodology that has been extensively applied in real-world applications to solve computer vision tasks. Nonetheless, the inherent challenges of deep learning models lie in their black-box nature, rendering them opaque and hard to interpret. Recently, attention-based vision transformers have been introduced to overcome the black-box behaviour of deep learning models. Despite these advances, the decision-making process of the vision transformer is still challenging to interpret. Learning classifier systems is a state-of-the-art rule-based evolutionary machine learning technique that stands out for its ability to provide interpretable decisions. These systems generate niche-based solutions, require less memory, and can be trained using small data sets. We hypothesize integrating attention mechanisms into learning classifier systems, aiming to identify critical components in problem instances, link features to create simple patterns, and model hierarchical relationships in the data. The experimental results for binary-class image classification (cat and dog) tasks demonstrate that the novel system successfully ignores the irrelevant parts and pays attention to the salient features of cats and dogs. Crucially, the proposed system exhibits comparable performance accuracy to that of the state-of-the-art learning classifier systems.

Accounting for wheel–terrain interaction is crucial for navigation and traction control of mobile robots in outdoor environments and rough terrains. Wheel slip is one of the surface hazards that needs to be detected to mitigate against the risk of losing the robot's controllability or mission failure occurring. The open problems in the Terramechanics field addressed are (1) the need for in situ wheel‐slippage estimation in harsh environments using low‐cost/power and easy to integrate sensors, and (2) removing the need for prior information of the soil, which is not always available. This paper presents a novel slip estimation method that utilizes only two proprioceptive sensors (IMU and wheel encoder) to estimate the wheel slip using deep learning methods. It is experimentally shown to be real‐world feasible in outdoor, uneven terrains without prior soil information assumptions. Comparison with previously used machine learning algorithms for continuous and discrete slip estimation problems show more than 9% and 14% improvement in estimation performance, respectively.

Deep-learning and large scale language-image training have produced image object detectors that generalise well to diverse environments and semantic classes. However, single-image object detectors trained on internet data are not optimally tailored for the embodied conditions inherent in robotics. Instead, robots must detect objects from complex multi-modal data streams involving depth, localisation and temporal correlation, a task termed embodied object detection. Paradigms such as Video Object Detection (VOD) and Semantic Mapping have been proposed to leverage such embodied data streams, but existing work fails to enhance performance using language-image training. In response, we investigate how an image object detector pre-trained using language-image data can be extended to perform embodied object detection. We propose a novel implicit object memory that uses projective geometry to aggregate the features of detected objects across long temporal horizons. The spatial and temporal information accumulated in memory is then used to enhance the image features of the base detector. When tested on embodied data streams sampled from diverse indoor scenes, our approach improves the base object detector by 3.09 mAP, outperforming alternative external memories designed for VOD and Semantic Mapping. Our method also shows a significant improvement of 16.90 mAP relative to baselines that perform embodied object detection without first training on language-image data, and is robust to sensor noise and domain shift experienced in real-world deployment.

Evolutionary Computation (EC) often throws away learned knowledge as it is reset for each new problem addressed. Conversely, humans can learn from small-scale problems, retain this knowledge (plus functionality) and then successfully reuse them in larger-scale and/or related problems. Linking solutions to problems together has been achieved through layered learning, where an experimenter sets a series of simpler related problems to solve a more complex task. Recent works on Learning Classifier Systems (LCSs) has shown that knowledge reuse through the adoption of Code Fragments, GP-like tree-based programs, is plausible. However, random reuse is inefficient. Thus, the research question is how LCS can adopt a layered-learning framework, such that increasingly complex problems can be solved efficiently? An LCS (named XCSCF*) has been developed to include the required base axioms necessary for learning, refined methods for transfer learning and learning recast as a decomposition into a series of subordinate problems. These subordinate problems can be set as a curriculum by a teacher, but this does not mean that an agent can learn from it. Especially if it only extracts over-fitted knowledge of each problem rather than the underlying scalable patterns and functions. Results show that from a conventional tabula rasa, with only a vague notion of what subordinate problems might be relevant, XCSCF* captures the general logic behind the tested domains and therefore can solve any n-bit Multiplexer, n-bit Carry-one, n-bit Majority-on, and n-bit Even-parity problems. This work demonstrates a step towards continual learning as learned knowledge is effectively reused in subsequent problems.

As machine learning models become increasingly prevalent in everyday life, there is a growing demand for explanation of the predictions generated by these models. However, most models used by companies are black-boxes in nature, without the capacity to provide explanations to users. This reduces public trust in these models, and exists as a barrier to adoption of machine learning. Research into providing explanations to users has shown that local explanation techniques provide more acceptable explanations to users than attempting to explain an entire model, as a user often does not need to understand the entirety of a model. This work builds on prior work in the field to produce a competitive method for high-fidelity local explanations utilising genetic programming. Two different data representations targeted towards both users with and without machine learning experience are evaluated. The experimental results show comparable fidelity to the state-of-the art, while exhibiting more comprehensible explanations due to including fewer features in each explanation. The method enables decomposable explanations that are easy to interpret, while still capturing non-linear relationships in the original model.

Perceptual aliasing is a cognitive problem for a learning agent where the robot cannot distinguish its state via its immediate observations, leading to poor decision-making. Previous work addresses this issue by storing the agent's path to learn the optimal policy. In particular, FoRsXCS utilises a fundamental and unique path to identify and disambiguate all aliased states and learn optimal policies in an environment with aliased states. However, it is hard to identify the aliased states in sequential aliasing patterns of states where the aliased states occur sequentially within a regular pattern. This work proposes a new cognitive learning system to identify such sequential aliasing patterns of states by extending FoRsXCS. The experimental results show that the proposed system performs equal to or greater than the existing systems in nine mazes for navigation tasks and significantly outperforms existing techniques in mazes with sequential aliasing patterns. Concretely, the proposed method improves its performance by 0.48 steps compared with FoRsXCS.

The abstraction and reasoning corpus (ARC) is a challenging AI benchmark as it requires models to learn unseen relationships from a few data points. Each puzzle only contains 2-5 training examples, which makes it hard for models that require training on large datasets. Models which do not require training on large datasets like learning classifier systems (LCSs) have the potential to solve this kind of complex, low-data problem due to their flexible representation, niche-based learning, and ability to generalise. Whilst some learners that can operate on low data have been applied to ARC, LCS-based architectures remain entirely unexplored. We propose a simple LCS architecture employing a windowing approach. This architecture solves 19 of 400 test grids (4.75%) in the ARC training set, which shows promise by outperforming other naive approaches. Additionally, the system uses minimal prior knowledge to achieve this result, bringing it closer to the original vision of an ARC solver, which is to only rely on a core set of concepts. We provide directions on how this basic model could be expanded upon to include more complex structures making use of LCSs' ability to integrate many diverse kinds of representations.

Unsupervised Domain Adaptive Object Detection (UDA-OD) uses unlabelled data to improve the reliability of robotic vision systems in open-world environments. Previous approaches to UDA-OD based on self-training have been effective in overcoming changes in the general appearance of images. However, shifts in a robot's deployment environment can also impact the likelihood that different objects will occur, termed class distribution shift. Motivated by this, we propose a framework for explicitly addressing class distribution shift to improve pseudo-label reliability in self-training. Our approach uses the domain invariance and contextual understanding of a pre-trained joint vision and language model to predict the class distribution of unlabelled data. By aligning the class distribution of pseudo-labels with this prediction, we provide weak supervision of pseudo-label accuracy. To further account for low quality pseudo-labels early in self-training, we propose an approach to dynamically adjust the number of pseudo-labels per image based on model confidence. Our method outperforms state-of-the-art approaches on several benchmarks, including a 4.7 mAP improvement when facing challenging class distribution shift.