Development Of an Algorithm for Strategic Group Allocation

Practice Paper Recommended citation: Guo, L., & Jiang, J. (2025). Development Of an Algorithm for Strategic Group Allocation. In Kangaslampi, R., Langie, G., Järvinen, H.-M., & Nagy, B. (Eds.), SEFI 53rd Annual Conference. European Society for Engineering Education (SEFI), Tampere, Finland. DOI: 10.5281/zenodo.17631842. This Conference Paper is brought to you for open access by the 53rd Annual Conference of the European Society for Engineering Education (SEFI) at Tampere University in Tampere, Finland. This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License. Development of an algorithm for strategic group allocation L. Guo a, 1 , J. Jiang b a Department of Mechanical Engineering, University College London, London, UK, 0000-0001-5576-4559 b Department of Mechanical Engineering, University College London, London, UK, 0009-0005-4331-3212 Conference Key Areas: 4. Digital tools and AI in engineering education; 10. Engineering skills, professional skills, and transversal skills. Keywords: Teamwork, group allocation, genetic algorithm, mechanical engineering, professional skill. ABSTRACT Teamwork is often highly rated by employers as one of the most important professional skills for engineering graduates. Group allocation has a significant influence on the group performance and student learning experiences. In this practice paper we present a new algorithm for strategic group allocation developed in the Department of Mechanical Engineering at University College London. The aim is to achieve balanced skill sets within each group so students can perform well collaboratively as a group no matter what the task is, and they not only learn from the teaching staff but also from each other. The algorithm starts by collecting student self-evaluation data from a skill survey and then uses a standard genetic algorithm to optimise group allocation. So far, the data shows students worked well in the groups allocated by the new algorithm, and teaching staff also praised the streamlined workflow of the algorithm. 1 Corresponding Author L. Guo [email protected].uk 1 INTRODUCTION Teamwork is an essential professional skill for engineering graduates. Creating effective teamwork environments at universities, therefore, is critical to prepare our students for the job market. The first step in teamwork is group allocation. Currently the most common methods for group allocation are random allocation and selfselection, where students can choose their teammates to form their own groups. Random allocation is an easy operation, but it ignores individual student’s strengths and weaknesses, whereas allowing students to form their own groups often ends up with students choosing their friends instead of considering the skills required for group tasks (Hanley, 2022). To overcome such limitations, strategic group allocation can achieve balanced skillsets and also take other factors into consideration, such as EDI (Equality, Diversity and Inclusion) and student preferences. However, the manual work is very time-consuming, especially for large cohorts and extensive skill evaluations. Therefore, there is a strong need to automate such processes using advanced computational algorithms. There are tools, such as CATME (Layton et al., 2010), available for tutor specified group allocation; however, many such tools focus on peer evaluation, such as SPARKPLUS (Willey and Gardner, 2013). Other previous work applied similar methods for group allocation, such as using psychological questionnaire (Wang et al., 2007) and grouping based on specific preferred skills for the tasks (Ani et al, 2010; Moreno et al, 2012; Nand et al., 2018). In this practice paper we present a new group allocation algorithm developed in the Department of Mechanical Engineering at University College London. We aim to develop an algorithm for automatic and strategic group allocation and focus on achieving balanced distribution of a wide range of skills. The allocation should have balanced skill sets within each group, so students can complement each other and perform well collaboratively as a group no matter what the task is. It can also promote a more efficient learning environment, where students not only learn from the teaching staff but also from each other. 2 CONTEXT AND PRACTICAL WORK 2.1 Overall structure The entire algorithm includes three steps: skill survey, data processing and group allocation. We have designed a customisable skill survey to get self-evaluated data from students. Then we feed the collected data into our group allocation software, which first performs data processing to obtain normalised scores and then allocate students into groups using a genetic algorithm. The overall structure of the algorithm is shown in Fig. 1, and each key step is future described in the following sections. Fig. 1. Flowchart of the overall structure Fig. 2. Flowchart of the genetic algorithm 2.2 Skill survey The group allocation is based on the data collected from a skill survey, where we ask students to self-evaluate their knowledge and skills in a comprehensive questionnaire (Table 1). More specifically, we consider three categories – engineering knowledge, technical skills and professional skills. As the current algorithm is developed for use in the Department of Mechanical Engineering, the engineering knowledge and technical skills focus more on this specific discipline; the professional skills, however, are transferred and applicable to other disciplines as well. Moreover, the next stage of algorithm development will focus on generalising the questionnaire and making the algorithm customisable for different disciplines. The advantage of asking students to self-evaluate their knowledge and skills is we can get a more complete picture of their skill sets, because other evaluation methods, e.g. exams, only cover certain modules and the data may even come from different sources, e.g. newly enrolled Master’s students from different universities. However, self-evaluation unavoidably introduces biases into the answers, especially in the category of professional skills, where students’ own perceptions can be subjective. Therefore, we normalise the raw data before feeding into the genetic algorithm for group allocation. For the categories of engineering knowledge and technical skills, we ask students to rate their proficiency on a scale of 1-5, whereas for the category of professional skills we ask students to give a ranking in each subcategory – communication, selfmanagement, project management and problem solving. The reason we use different evaluation methods is professional skills are difficult to quantify so a ranking is more accurate to reveal students’ relative strengths and weaknesses. We then use the following criterion to convert ranking into scores: 𝑅𝑎𝑛𝑘 1 → 𝑆𝑐𝑜𝑟𝑒 4 𝑅𝑎𝑛𝑘 2 → 𝑆𝑐𝑜𝑟𝑒 3 𝑅𝑎𝑛𝑘 3 → 𝑆𝑐𝑜𝑟𝑒 2 (1) A min-max normalisation is used to normalise the raw scores in the categories of engineering knowledge and technical skills. Min-max normalisation can rescale the dataset in a specific range; in this case, the scale is 1-5. This method is used because it can adjust the scores if a student only uses a part of the scale (for example, if they only rate themselves between 2 and 4) without distorting the relative differences between the data points so all students can be compared in the same way. The normalised score 𝑋′ is calculated from raw scores 𝑋 as 𝑋′=𝑋−𝑋𝑚𝑖𝑛 𝑋𝑚𝑎𝑥−𝑋𝑚𝑖𝑛 ×(𝑌 𝑚𝑎𝑥 − 𝑌 𝑚𝑖𝑛)+ 𝑌 𝑚𝑖𝑛 (2) where 𝑋𝑚𝑖𝑛 is the original minimum score given by a student; 𝑋𝑚𝑎𝑥 is the original maximum score given by a student; 𝑌 𝑚𝑖𝑛 is the desired minimum score (in this case 𝑌 𝑚𝑖𝑛 = 1); 𝑌 𝑚𝑎𝑥 is the desired maximum score (in this case 𝑌 𝑚𝑎𝑥 = 5). When all the original scores are the same, i.e. 𝑋𝑚𝑖𝑛 = 𝑋𝑚𝑎𝑥, 𝑋′=𝑌 𝑚𝑎𝑥+𝑌𝑚𝑖𝑛 2 (3) Table 1. Knowledge and skills evaluated in the skill survey Category Knowledge and skills Evaluation method Engineering knowledge Materials Scale of 1-5 Fluid mechanics Thermodynamics Solids and structures Mathematics Control & instrumentation Technical skills Computer-Aided Design (CAD) Finite Element Analysis (FEA) Computational Fluid Dynamics (CFD) Programming (Python/MATLAB) 3D printing CNC machining Professional skills Communication Verbal communication Ranking in each subcategory Written communication Presentation Self-management Time management Stress management Adaptability Leadership Project management Budgeting and business Planning and organisation Problem solving Practical skills Analytical thinking Creativity 2.3 Constraints in the allocation At the moment we consider two constraints in the group allocation: gender and native language. The criterion for gender constraint is to avoid allocating only one student of a certain gender into one group. The consideration here is gender minority in a group may feel isolated and uncomfortable working with others. Gender constraint is implemented when first initialising the population in the genetic algorithm so it means that during the iterations gender constraint cannot be violated. Similarly, native language constraint is introduced to create a diverse and multicultural learning environment. It is worth noting that there are some exceptional cases where it is not possible to implement such constraints. For example, in small groups of two students any composition is homogeneous (i.e. no minority); in groups of three students there is inevitable minority for one student versus the other two students unless the group composition is homogeneous (e.g. all members of the same gender). In such cases, gender and native language constraints are relaxed to only focus on achieving balanced skill sets. 2.4 Genetic algorithm Genetic algorithm (Holland, 1992), which is designed to mimic the mechanism of natural selection, is well-developed and widely used for optimisation problems (Nammuni et al., 2002). Here we use a standard genetic algorithm, aiming to minimise the variance of skill scores between groups to achieve balanced skill sets within each group. The flowchart of the key steps is shown in Fig. 2 and further explained below. The iterations will continue until the user specified iteration (generation) number is reached. Then the optimal allocation (i.e. the one with the lowest fitness score) is returned to the user. 1. Each population represents a specific configuration of group compositions. The first step initialises all populations based on the user specified value (e.g. 50). 2. A fitness score is calculated for each population, which considers the constraints applied (e.g. gender and native language) as well as the variance between groups, which is denoted by V and calculated as 𝑉 = ∑(𝑥𝑖−𝜇)2 𝑁 𝑖=1 𝑁 (4) where 𝑥𝑖 is the average score of group i; 𝜇 is the average value of all group average scores; 𝑁 is the number of groups. The aim here is to minimise the variance to achieve balanced skill sets within each group. Violation of constraints will add a penalty. Therefore, the lower the fitness score, the better. 3. New configurations are created by crossover, which means swapping members between groups. 4. Mutation is applied to maintain the diversity in group compositions, i.e. preventing configurations becoming too similar. 3 RESULTS AND INSIGHTS The algorithm is coded in Python. Currently we are doing a trial of the group allocation algorithm for two Master’s level modules in our department – both involving group-based projects. Here we present the preliminary data collected so far. One of the modules in the trial (MSc Group Design Project) has a cohort of about 100 students. When we used the algorithm to allocate students into groups of eight, the fitness score showed good convergence (Fig. 3). Fig. 3. Convergence of the fitness score with respect to the generation number Another module in the trial (Group Manufacturing Challenges) has a small cohort of nine students, who were allocated into three groups of three. As described in Section 2.3, in this case the gender and native language constraints are not applicable anymore due to the small group size. We continuously monitored peer assessment scores, where students evaluate the performance for their teammates and themselves. A score above 1.0 means the student is outperforming their teammates, whereas a score below 1.0 means the student is underperforming (del Pilar Garcia Souto et al., 2019). In theory, every member achieving a score of 1.0 is an indicator of good group dynamics – meaning all members work in a collaborative and harmonious way. The range of peer assessment scores within each group is calculated as 𝑚𝑎𝑥 − 𝑚𝑖𝑛 and plotted in Fig. 4 for all the peer assessments conducted in Term 1 (October-December 2024). At the beginning group members had more divergent performances and then gradually converged towards more equal contributions, which implies that the group allocation worked well to improve productivity and collaboration. This is also reflected in the comments given by students in the peer assessments, some of which are extracted below. Early stage: “… less willing to interact and communicate with us”, “… perhaps communication and more active participation can be improved upon”, “… need to be more proactive in sharing opinions and asking questions”. Late stage: “Excellent teamwork”, “Proactive and effective communication”, “Communication was kept clear and we were able to work well together to produce a great design”. Fig. 4. Range of peer assessment scores in the trial We also conducted a follow-up survey to evaluate skill developments of students after group projects for the module with nine students. We asked students to choose one of the quantitative indicators (<0%, 0%, 20%, 50%, >80%) to measure their developments of different skills. We gave students the options of zero and even negative percentages if they feel they did not achieve any improvements at all or even worse. All nine students completed the survey and the results are shown in Fig. 5, where each column shows the average value of percentage scores of the specific skill and each error bar shows the maximum and minimum scores received for that skill. The skills are categorised into three groups – engineering knowledge and technical skills (blue colour), general professional skills (orange colour) and more detailed teamwork skills (green colour). It is worth noting that the skills covered in this follow-up survey is less than the ones in the initial survey for group allocation (Table 1), because not all skills are learning outcomes of this specific module. It can be seen that students achieved improvements in all skills, with the highest average score of 73% in practical skills, which fits the learning outcomes of this module very well. In this module, students work in small groups to apply the knowledge they learnt from other modules in the MSc programme to tackle realworld problems, so practical skills are among the most important skills we hope they can develop throughout the projects. It is also interesting to see that students generally developed better in technical skills than professional and teamwork skills, not only reflected by the higher average scores but also the distribution, especially the minimum scores. Every skill has some students choosing the highest 80% but only the technical skills (blue colour) have no scores below 20%. This can be interpreted from at least two angles. First, it is much more intricate to “teach” soft skills, which is affected by far more factors than academic performance; Second, the self-evaluation of professional skills in the initial survey may be too subjective and therefore not very accurate to allocate students into the best matched groups to develop such skills. Fig. 5. Results of a follow-up skill survey after group projects 4 CONCLUSIONS AND IMPLICATIONS We have developed an algorithm for strategic group allocation, aiming to achieve balanced skill sets within each group. The algorithm starts by collecting student selfevaluation data from a skill survey and then uses a standard genetic algorithm to optimise group allocation. So far, the preliminary data shows students worked well in the groups allocated by the new algorithm. For example, students explicitly mentioned the positive effects of different roles working collectively in the group. Teaching staff also praised the streamlined workflow of the algorithm. This practice paper summarises our work in progress and presents the algorithm as a technical tool, which can be readily used by Mechanical Engineering departments at other universities. However, more work is needed to further develop it into an advancement in educational methodology. For example, future work includes adding customised functions to accommodate different needs of other disciplines, and adding more source data for group allocation, e.g. Belbin test, to more accurately reflect the soft skills. In terms of trials, comparisons with control cases (i.e. groups not formed using this algorithm) are needed to assess its impact. 5 ACKNOWLEDGEMENTS This project was supported by studentships funded by the Department of Mechanical Engineering and the Centre for Engineering Education at University College London.