6G wireless networks are envisioned to deliver pervasive, intelligent, and resilient connectivity across terrestrial, aerial, and space domains, supporting ultra-high data rates, sub-millisecond latency, massive device density, and tightly integrated communication, computation, and sensing. Realising this vision requires adaptive network intelligence capable of operating under uncertainty in highly dynamic environments, motivating the adoption of computational intelligence (CI), which brings together machine learning and optimisation for continual adaptation, autonomous decision-making, and robust control. Although a growing body of survey literature examines AI in 6G, most existing works focus on isolated technologies, application domains, or specific model classes, providing limited guidance on which CI methods are structurally suited to particular 6G tasks. This distinction matters because deep reinforcement learning, evolutionary algorithms, federated learning, graph neural networks, and fuzzy systems exhibit different inductive biases whose effectiveness depends on the underlying problem structure. This survey addresses this gap through a method-centric synthesis of CI across nine representative 6G enabling technologies: THz communications, reconfigurable intelligent surfaces, integrated sensing and communication, AI-native communications, extreme edge computing, zero-energy communications, space-air-ground integrated networks, semantic communications, and digital twins. For each technology, canonical tasks are identified, representative CI families are mapped to those tasks, and methodological trade-offs are analysed. The survey introduces the method–structure alignment principle as a unifying lens for CI method selection, identifies recurring conditions under which methods succeed or fail, and presents a consolidated research agenda for robust, scalable, cross-domain, and responsible CI in 6G.

Neural architecture search explores the architecture space, referred to as the design spaces, to find an architecture that produces good results. Various approaches, such as genetic algorithms, are usually used to explore this space. This study investigates exploring an alternative space, namely, the heuristic space using a hyper-heuristic to indirectly explore the design space. The study introduces the concept of a NAS operator space (NOS). A single point selection perturbative hyper-heuristic(SPHH-NAS) explores a heuristic space that maps to the NOS which then maps to the design space. A choice function is used for heuristic selection and the Adaptive Improvement Limited Target Acceptance (AILTA) for move acceptance. It is anticipated that indirectly searching the design space will facilitate reaching areas of the search space that could not be reached by searching the space directly. SPHH-NAS was evaluated on three NAS benchmark sets, namely, NAS-101, NAS-201 and NAS-301. In addition to this the approach is evaluated on two real-world datasets. SPHH-NAS was found to outperform majority of the previous approaches used to solve these problems. In addition to this SPHH-NAS resulted in a reduction in computational cost.

Critical infrastructure systems (CISs) are the backbone of every country and the world as a whole, with the ultimate goal of maintaining resilient CISs at a reasonable cost. This position paper claims that artificial intelligence (AI) is essential for CISs. Critical infrastructure sectors common to most countries include communication systems, energy systems, transportation systems, water systems, agriculture, healthcare and finance. The paper firstly shows how AI has been used in CISs in each of these sectors. We refer to this as AI for critical infrastructure (AI4CI). The use of AI in these sectors is presented in terms of the basic functionality of the CISs, fault prediction in CISs, monitoring and maintenance by CISs and security. While there are various challenges associated with AI4CI, such as the need for explainability, computational cost, data requirements and associated risks, e.g., biases and adversarial attacks, the paper argues that methods to overcome these, including AI approaches, have been identified and have proven to be effective. Natural disasters, terrorism, cyber threats, aging and decay of infrastructure, water contamination and cascading CIS failures threaten CIS resilience. The paper illustrates how AI has been used to predict and eliminate these threats. This paper has shown that AI plays a critical role in CISs and will continue to do so. The research done in this field has identified further areas of research needed in AI4CI, which the paper also highlights.

In previous work, algorithms have been decomposed into basic algorithmic components and then recomposed into brand new algorithms using a genetic algorithm. The algorithm is composed in full prior to execution of the algorithm. We refer to this as static algorithm composition (SAC). This study examines composing segmentation algorithms in real-time by varying the techniques used for each of the algorithmic components as well as the order of the algorithmic components at different points during the execution of the segmentation algorithm. A genetic algorithm (GA) is used to compose the segmentation algorithm. Furthermore, the techniques that can be used for each algorithmic component changes every$g$generations of the GA. A selection perturbative hyper-heuristic is used to determine the techniques that the GA can use for each of the algorithmic components. We refer to this as dynamic algorithm composition (DAC). Static and dynamic algorithm composition is evaluated for image segmentation using the BSD-500, PASCAL VOC, Football and COVID CT scan datasets. DAC improved on SAC for all datasets. Additionally, DAC was found to be competitive with the state of the art and improved on the best known results for the COVID CT scan datasets.

Genetic programming and its variants, such as grammatical evolution, have been predominantly used for generating functions for machine learning techniques, such as loss or activation functions for neural networks and choice functions for the Fuzzy ART algorithm. These functions are evolved offline prior to the execution of the neural network and remains the same during the execution of the learning algorithm. We refer to this as static function generation (SFG). This study examines generating these functions in real-time at different points during the execution of the machine learning algorithm. We refer to this as dynamic function generation (DFG). Grammatical evolution (GE) is used to generate the function. Each function is generated at every$m$epochs of the learning algorithm. Furthermore, the grammar used by GE also changes every$g$generations of the GE algorithm. A selection perturbative hyper-heuristic is used to determine the options to include in the grammar. In previous work the effectiveness of using GE to evolve the choice function for the Fuzzy Art algorithm was shown. We use this as a case study to investigate DFG given the success of generating choice functions for this learning algorithm in previous work. However, DFG can be used with any neural network learning algorithm. Static and dynamic function generation is evaluated for text classification using the Enron, SMS Spam, Chat GPT tweets, IMDB movie reviews and Amazon product reviews datasets. DFG improved on the performance of SFG for all datasets. Additionally, DFG was found to be competitive with the state of the art and improved on the best known results for the SMS Spam and Chat GPT Tweets datasets.

Heat sinks are a prevalent and direct solution for addressing the Microelectronic Thermal Management Problem (MTMP), which is critical in today's electronic industry. Specifi-cally, an optimally designed thermodynamic heat sink ensures that microelectronics operate reliably without compromising their lifespan and performance, thereby indirectly safeguarding user safety. Although Metaheuristics (MHs) have proven effective in tackling this complex design challenge due to their robust characteristics, no single MH consistently delivers superior per-formance across all scenarios. The study explores the feasibility of an Automated Metaheuristic Design strategy, employing a hyper-heuristic search to develop a population-based, metaphor-free MH specifically for the MTMP. Various scenarios are assessed by varying the heat sink design specifications and benchmarking the custom MH designs against several state-of-the-art MHs. The findings of this preliminary work provide statistical evidence that the tailored MHs surpass the performance of established MHs in these scenarios. A toolkit of MH components is assembled, which can be customized to construct MHs specifically for MTMPs. This approach enables practitioners to select the most suitable solver for a particular problem without needing extensive expertise in heuristic-based optimization.

Recent work has shown the potential of ant algorithms for generation constructive hyper-heuristics. This paper extends the previous research by presenting a novel ant algorithm that is used to drive the heuristic search for a generation perturbative hyper-heuristic, the other type of generation hyper-heuristic. The ant-based generation perturbative hyper-heuristic is presented and compared against existing heuristics in two combinatorial domains, the movie scene scheduling and capacitated vehicle routing problems, to assess the heuristic generation efficacy. The comparison is further extended by assessing the effect of different pheromone maps (1D, 2D and 3D) on the ant-based hyper-heuristic, an important factor in the previous study. The results showed that, in both domains, the hyper-heuristic was able to generate perturbative heuristics that were competitive or better than the existing heuristics. Furthermore, the type of pheromone map was relevant to the hyper-heuristic performance as the 3D map performed best for the first domain and the 1D map for the second, confirming the trend shown in previous research, as well as the validity of ant algorithms for generation perturbative hyper-heuristics.

It is no secret that optimisation is a popular topic in any practical engineering application. Similarly, Metaheuristics (MHs) are a fairly standard approach for solving optimisation problems due to their success, flexibility, and simplicity. However, it is seldom easy to find a solver from the overpopulation of metaheuristics that adequately deals with a given problem. For that reason, the solver selection is even considered an additional problem in many optimisation scenarios. This work investigates the Metaheuristic Composition Optimisation Problem, which involves designing heuristic-based procedures that solve continuous optimisation problems. Therefore, we propose two novel and still simple methodologies based on transfer learning to facilitate the automatic generation of population-based and metaphor-less MHs by using search operators from the literature. To represent these solvers, we adopt our previously proposed unfolded MH model. The first strategy deals with the problem dynamically, building the sequence while solving the low-level problem. In contrast, the second one does it statically by generating the whole candidate sequence before implementing it. Results provide us with information to prove the feasibility of these approaches via experiments using 32 problems with four different characteristic groups and four dimensionalities and varying the number of agents (30, 50, and 100) employed by the search operators. We also remark that one can compare these two methodologies on performance, but we emphasise their potential usage depending on the general application environment.

The automated design of search techniques is a recent trend in artificial intelligence research. Unfortunately, the majority of the automated design approaches are developed using trial and error which fails to justify or at least explain why some design decisions succeed while others fail. This approach is a host of evils as it has resulted in poorly understood systems for poorly understood problems. This study is an attempt to improve our understanding of the automated design of hybrid metaheuristics by utilizing fitness landscape analysis to reveal the topological characteristics that can be exploited to design better automated approaches. We consider the sequential hybridization, including algorithm configuration and parameter tuning, of single-point and multi-point metaheuristics and three optimization problems which are the earth-observing satellite scheduling problem, the aircraft landing problem and the two-dimensional bin packing problem. Interestingly, the design space exhibits similar trends regardless of the underlining optimization problem. The design space is found to be rugged, multimodal, moderately searchable, has multiple funnels, and almost no plateau. Based on these findings, deeper insights are provided to guide the development of future automated approaches instead of blindly trying different options.

The automated design of optimization techniques holds great promise for advancing state-of-the-art optimization techniques and it has already taken over the manual design by human experts in some problems. Genetic algorithms are one of the key approaches for tackling the automated design problem. Unfortunately, these algorithms may take several hours to run as the fitness evaluation involves solving some benchmark instances to determine the quality of a candidate configuration. In this paper, we hybridize a meta-genetic algorithm with reinforcement learning to automatically design genetic algorithms for the two-dimensional bin packing problem. The task of the meta-genetic algorithm is to search the configuration space of genetic algorithms and the task of reinforcement learning is to decide whether to evaluate a candidate configuration or not. Therefore, avoiding wasting the computational budget on poor configurations. The proposed hybrid and the meta-genetic algorithm without reinforcement learning produce solvers for the two-dimensional bin packing problem that are competitive with the state-of-the-art algorithms. However, the proposed hybrid consumes about 25% of the computational effort required by the meta-genetic algorithm without reinforcement learning.

The travelling salesperson problem is an NP-hard combinatorial optimization problem. In this paper, we consider the multi-objective travelling salesperson problem (MTSP), both static and dynamic, with conflicting objectives. NSGA-II and MOEA/D, two popular evolutionary multi-objective optimization algorithms suffer from loss of diversity and poor convergence when applied separately on MTSP. However, both these techniques have their individual strengths. NSGA-II maintains di-versity through non-dominated sorting and crowding distance selection. MOEA/D is good at exploring extreme points on the Pareto front with faster convergence. In this paper, we adopt the bicriterion framework that exploits the strengths of Pareto-Criterion (PC) and Non-Pareto Criterion (NPC) evolutionary populations. In this research, NSGA-II (PC) and MOEA/D (NPC) coevolve to compensate the diversity of each other. We further improve the convergence using local search and a hybrid of order crossover and inver-over operators. To our knowledge, this is the first work that combines NSGA-II and MOEA/D in a bicriterion framework for solving MTSP, both static and dynamic. We perform various experiments on different MTSP bench-mark datasets with and without traffic factors to study static and dynamic MTSP. Our proposed algorithm is compared against standard algorithms such as NSGA-II & III, MOEA/D, and a baseline divide and conquer coevolution technique using performance metrics such as inverted generational distance, hypervolume, and the spacing metric to concurrently quantify the convergence and diversity of our proposed algorithm. We also compare our results to datasets used in the literature and show that our proposed algorithm performs empirically better than compared algorithms.

Generation construction hyper-heuristics have proven to be effective in solving discrete optimization problems. Previous work has shown the effectiveness of an ant colony optimization hyper-heuristic for solving scheduling and packing problems. One of the challenges with generation construction hyper-heuristics is the high processing times associated with creating new construction heuristics. While there has been research into using transfer learning to reduce the computational cost of genetic programming generation constructive hyper-heuristics, this has not been investigated for ant colony optimization generation construction hyper-heuristics. In fact to the knowledge of the authors transfer learning has not previously been investigated for ant colony optimization. In this study the knowledge transferred is the pheromone map. The maps are transferred from the source domain to the target domain, with the target domain being more complicated problem instances and the source domain simpler problem instances, which do not take as long to solve. The approach was evaluated on the movie scene scheduling problem, the one dimensional bin packing problem and the quadratic assignment problem. The study has shown that the use of transfer learning has reduced the computational cost drastically while maintaining the same performance for the more complex problems for the movie scene scheduling problem and the quadratic assignment problem. However, for the one dimensional bin packing problem while there is a reduction in computational cost, the quality of the solutions is worse. Future research will investigate the reason for this and evaluate transferring different types of knowledge at various points in the life cycle of ant colony optimization generation construction hyper-heuristics.

Metaheuristics (MHs) are proven powerful algorithms for solving non-linear optimisation problems over discrete, continuous, or mixed domains. Applications have ranged from basic sciences to applied technologies. Nowadays, the literature contains plenty of MHs based on exceptional ideas, but often, they are just recombining elements from other techniques. An alternative approach is to follow a standard model that customises population-based MHs, utilising simple heuristics extracted from well-known MHs. Different approaches have explored the combination of such simple heuristics, generating excellent results compared to the generic MHs. Nevertheless, they present limitations due to the nature of the metaheuristic used to study the heuristic space. This work investigates a field of action for implementing a model that takes advantage of previously modified MHs by learning how to boost the performance of the tailoring process. Following this reasoning, we propose a hyper-heuristic model based on Artificial Neural Networks (ANNs) trained with processed sequences of heuristics to identify patterns that one can use to generate better MHs. We prove the feasibility of this model by comparing the results against generic MHs and other approaches that tailor unfolded MHs. Our results evidenced that the proposed model outperformed an average of 84 % of all scenarios; in particular, 89 % of basic and 77 % of unfolded approaches. Plus, we highlight the configurable capability of the proposed model, as it shows to be exceptionally versatile in regards to the computational budget, generating good results even with limited resources.

Metaheuristics are a fairly standard approach for solving optimisation problems due to their success, flexibility, and simplicity. However, there is a plethora of metaheuristics available, with different performance levels for various problems. This work proposes a methodology for designing heuristic-based procedures to solve continuous optimisation problems and study how the population size affects its performance. The technique comprises the well-known Simulated Annealing algorithm as a hyper-heuristic, and a heuristic sequence taken from unfolding the conventional scheme of population-based metaheuristics reported in the literature. Our results show that the proposed approach is a reliable alternative for tackling optimisation problems. We find exciting insights, according to our data, about this primary implementation when varying the population size in different challenging problems.

This study builds on previous work in which Mask R-CNN was used to quantify common rust severity on maize leaves grown in a greenhouse environment. The Mask R-CNN outperformed standard image processing techniques when applied to the rust quantification task. One of the challenges with using Mask R-CNN is the large amount of parameters that need to be tuned, this study examines automated tuning of these parameters. This study is the first to perform hyper-parameter tuning on the Mask R-CNN at this scale as well as using a genetic algorithm (GA). The results gathered from the GA implementation is compared to that of the prior results from the Mask R-CNN with manually selected hyper-parameters. This work concludes that using a genetic algorithm to determine optimal hyper-parameter values for the Mask R-CNN is comparable to the current state of the art for this problem domain. Future work includes performing a fitness landscape analysis and potential development of new genetic operators.

This special section is formed by 15 articles of outstanding quality that together comprise a snapshot of cutting edge automated machine learning (AutoML) research.

As we move into the fourth industrial revolution video streaming platforms like Netflix are turning to machine learning techniques to maintain a competitive edge in the market. Various problems such as clip creation, network optimization, customer churn prediction, amongst others, have been solved for video streaming platforms using machine learning. This paper focuses on automatic thumbnail selection for movies and series. Classifiers are used to automate the thumbnail selection. The research firstly compares the performance of different convolutional neural networks (CNNs), namely, VGG-19, Inception-v3 and ResNet-50, for solving this problem. The performance of two classifiers, namely, the best performing convolutional neural network and a hybrid approach combining a CNN and genetic programming, are compared for thumbnail selection. The CNN is used for feature extraction and genetic programming for classification. The ResNet-50 CNN outperformed the other CNNs. Both classifiers were successful for thumbnail selection with the convolutional neural network outperforming the hybrid classifier.

Perturbative heuristics, or move operators, are domain specific operators generally used with search techniques, for example the 2-opt operator for the travelling salesman problem. These operators are usually derived manually which is an extremely time consuming task. The study presented in this paper investigates automating this process focussing on deriving perturbative heuristics for discrete optimization problems. This research forms part of a larger initiative aimed at automating the design of machine learning and search techniques. While there has been some research on the automated generation of perturbative operators, this has not been a well researched domain and most work has focussed on recombining existing human derived perturbative heuristics or components thereof together with move acceptance criteria rather than producing new perturbative heuristics from scratch. In this research pertubative heuristics are defined in terms of basic actions and solution components. Grammatical evolution is used to combine the basic actions and solution components into perturbative heuristics. The performance of the generated heuristics is compared to human derived heuristics for two problems, namely, examination timetabling (ITC 2007 benchmark instances) and capacitated vehicle routing (Christofides and Golden instances). For both problems the generated perturbative heuristics outperformed the human derived heuristics. Given the potential of the automated generation of perturbative heuristics established in this study, future work will extend the this work by including conditional and iterative constructs in the perturbative heuristics as well as applying the approach to other problems.

The examination timetabling problem is a well researched problem. Various techniques have been applied to solving this NP-hard problem including genetic algorithms and hyper-heuristics. The developmental approach was created specifically for solving the examination timetabling problem. Studies on the developmental approach revealed the effectiveness of searching in the partial solution space, incrementally building solutions, in solving the examination timetabling problem. Research into solving the examination timetabling problem has also revealed that one of the challenges is that timetables of very different structure often have the same behaviour and hence objective value. Taking both these aspects into consideration this paper presents a structure-based partial solution search (SBPSS) to solve the examination timetabling problem. This multipoint search works in the partial solution space of timetables, incrementally building timetable solutions. Furthermore, while movement through the search space in traditional search is based on behaviour in terms of fitness or the objective value, SBPSS directs the search based on genotype in terms of the structure of the timetables as well as behaviour. The SBPSS was applied to the ITC 2007 examination timetabling track benchmark problem set from the second international timetabling competition. The performance of the SBPSS was found to be competitive to other state of the art approaches, producing the best result for five of the examination timetabling problem instances.

Previous work has established the effectiveness of meta-genetic algorithms for the automated design of hybrid metaheuristics. The design decisions made by the meta-genetic algorithm include which metaheuristics to combine, the order of the metaheuristics in the combination and the parameter values to use for each metaheuristic. This research is still in its infancy and the research presented in this paper aims to improve on this initial study. Two areas for improvement in applying meta-genetic algorithms for hybridizing metaheuristics in this way have been identified. The first is to reduce the parameter space explored by the meta-genetic algorithm. Orthogonal arrays are examined for this purpose. The second improvement is in terms of the search by improving exploitation of the meta-genetic algorithm. An extension of the fitness-based scanning crossover is investigated as an alternative to one-point crossover to achieve this. The aircraft landing problem (ALP) is used to test these improvements. The evolved hybrid metaheuristics are found to perform competitively with the state-of-the-art methods and outperform the automatically tuned metaheuristics when used individually. The evolved hybrid metaheuristics demonstrate the reusability of the meta-genetic algorithm as they are designed using a relatively small training set and generalize to the whole dataset. In future work, other mechanisms for hybridizing metaheuristics shall be considered.

Hyper-heuristics is an emergent technology that has proven to be effective at solving real-world problems. The two main categories of hyper-heuristics are selection and generation. Selection hyper-heuristics select existing low-level heuristics while generation hyper-heuristics create new heuristics. At the inception of the field single point searches were essentially employed by selection hyper-heuristics, however as the field progressed evolutionary algorithms are becoming more prominent. Evolutionary algorithms, namely, genetic programming, have chiefly been used for generation hyper-heuristics. Implementing evolutionary algorithm hyper-heuristics can be quite a time-consuming task which is daunting for first time researchers and practitioners who want to rather focus on the application domain the hyper-heuristic will be applied to which can be quite complex. This paper presents a Java toolkit for the implementation of evolutionary algorithm hyper-heuristics, namely, EvoHyp. EvoHyp includes libraries for a genetic algorithm selection hyper-heuristic (GenAlg), a genetic programming generation hyper-heuristic (GenProg), a distributed version of GenAlg (DistrGenAlg) and a distributed version of GenProg (DistrGenProg). The paper describes the libraries and illustrates how they can be used. The ultimate aim is to provide a toolkit which a non-expert in evolutionary algorithm hyper-heuristics can use. The paper concludes with an overview of future extensions of the toolkit.

In solving combinatorial optimization problems construction heuristics are generally used to create an initial solution which is improved using optimization techniques like genetic algorithms. These construction heuristics are usually derived by humans and this is usually quite a time consuming task. Furthermore, according to the no free lunch theorem different heuristics are effective for different problem instances. Ideally we would like to derive construction heuristics for different problem instances or classes of problems. However, due to the time it takes to manually derive construction heuristics it is generally not feasible to induce problem instance specific heuristics. The research presented in the paper forms part of the initiative aimed at automating the derivation of construction heuristics. Genetic programming is used to evolve construction heuristics for the curriculum based university course timetabling (CB-CTT) problem. Each heuristic is a hierarchical combination of problem characteristics and a period selection heuristic. The paper firstly presents and analyses the performance of known construction heuristics for CB-CTT. The analysis has shown that different heuristics are effective for different problem instances. The paper then presents the genetic programming approach for the automated induction of construction heuristics for the CB-CTT problem and evaluates the approach on the ITC 2007 problem instances for the second international timetabling competition. The evolved heuristics performed better than the known construction heuristics, producing timetables with lower soft constraint costs.

There has been a fair amount of research into the use of genetic programming for the induction of game playing strategies for board games such as chess, checkers, backgammon and Othello. A majority of this research has focused on developing evaluation functions for use with standard game playing algorithms such as the alpha-beta algorithm or Monte Carlo tree search. The research presented in this paper proposes a different approach based on heuristics. Genetic programming is used to evolve game playing strategies composed of heuristics. Each evolved strategy represents a player. While in previous work the game playing strategies are generally created offline, in this research learning and generation of the strategies takes place online, in real time. An initial population of players created using the ramped half-and-half method is iteratively refined using reproduction, mutation and crossover. Tournament selection is used to choose parents. The board game Othello, also known as Reversi, is used to illustrate and evaluate this novel approach. The evolved players were evaluated against human players, Othello WZebra, AI Factory Reversi and Math is fun Reversi. This study has revealed the potential of the proposed novel approach for evolving game playing strategies for board games. It has also identified areas for improvement and based on this future work will investigate mechanisms for incorporating mobility into the evolved players.

This manuscript proposes a hyper-heuristic approach towards mitigating Premature Convergence caused by objective fitness in Genetic Programming (GP). The objective fitness function used in standard GP has the potential to profoundly exacerbate Premature Convergence in the algorithm. Accordingly several alternative fitness measures have been proposed in GP literature. These alternative fitness measures replace the objective function, with the specific aim of mitigating this type of Premature Convergence. However each alternative fitness measure is found to have its own intrinsic limitations. To this end the proposed approach automates the selection of distinct fitness measures during the progression of GP. The power of this methodology lies in the ability to compensate for the weaknesses of each fitness measure by automating the selection of the best alternative fitness measure. Our hyper-heuristic approach is found to achieve generality in the alleviation of Premature Convergence caused by objective fitness. Vitally the approach is unprecedented and highlights a new paradigm in the design of GP systems.

The study presented in this paper forms part of a larger initiative aimed at creating a generic architecture for the development of intelligent programming tutors (IPTs) in an attempt to reduce the costs associated with building IPTs. Thus, instead of requiring the lecturer to provide solution algorithms to the programming problems that students will be tested on by the system, the generic architecture will automatically generate the solutions to these problems. This paper reports on the results of an investigation conducted to test the hypothesis that genetic programming (GP) can be used for this purpose. The paper proposes a genetic programming system for the induction of solutions to arithmetic, character and string manipulation, conditional, iterative, nested iteration, and recursive problems. The paper analyses the results of applying the proposed system to 45 randomly chosen novice procedural programming problems. Extensions made to the proposed system based on this analysis, namely, the implementation of the iterative structure-based algorithm (ISBA), are discussed.

Spreadsheets have become a routine application in most organizations and universities. As a consequence, students are required to learn spreadsheet applications such as Microsoft Excel. The learning of spreadsheets is often accompanied by problems related to spreadsheet application and their mathematical content. The EXITS (Excel intelligent tutoring system) research project aims to develop a Microsoft Excel tutor that help students or learners to overcome their learning difficulties. In this paper, we analyse and classify spreadsheet errors made by students in order to determine the function that our system should perform and to generate an error library for student modelling purposes.