Machine Learning and Large Language Models for Automating ICT Installation Planning : A Comparative Study of Dependency Prediction Methods

Sammanfattning

Information and Communication Technology (ICT) installation planning in laboratory and datacenter environments is a complex, predominantly manual process that depends on expert knowledge and fragmented, semi-structured data sources. This thesis investigates whether dependency prediction in ICT installation planning can be learned from historical data, and compares three different modeling approaches including deterministic, machine learning (ML), and large language model (LLM). The study is based on historical installation planning data extracted from spreadsheet records and Configuration Management Database (CMDB) systems. These data represent structured engineering decisions describing physical connectivity between hardware components, including port assignments, cable types, and endpoint mappings. A key challenge is that the data are semi-structured, incomplete, and heterogeneous, raising questions about their suitability for data-driven learning. Three modeling approaches are evaluated. First, a deterministic Knowledge-Based Agent (KBA) is developed using explicitly encoded engineering constraints. Second, three supervised ML models (CatBoost, LightGBM, and XGBoost) are trained to learn planning decisions from historical data. Third, three instruction-tuned LLMs (FLAN-T5 Small, Qwen2.5-0.5B-Instruct, and Llama-3.2-1B-Instruct) are fine-tuned using structured input-output representations to perform the same prediction task. Model performance is evaluated using a dual-level framework, including both per-target classification metrics (accuracy and weighted F1-score) and work-order-level exact match accuracy. The results demonstrate that both ML and LLM approaches are capable of learning meaningful patterns from historical data, but their performance is limited by data sparsity, label imbalance, and inconsistencies in the underlying records. The Knowledge-Based Agent provides stable and interpretable output, highlighting the importance of domain constraints in this setting. The findings indicate that while data-driven models can support dependency prediction, their effectiveness is strongly dependent on data quality. The thesis concludes that hybrid approaches combining deterministic rules with data-driven methods are a promising direction for supporting ICT installation planning, rather than fully replacing existing workflows.