AI Institutes Seminar Date 25.5.26 Monday

Speaker: Lior Siag Supervisor: Prof. Ariel Felner and Dr. Shahaf Shperberg Date: Monday, 25/5, talk starts at 12:05 Place: Building 37, room 202 Talk title: New Directions in Bidirectional Search Abstract: Heuristic search uses heuristic functions to efficiently explore complex search spaces. Bidirectional heuristic search (BiHS) extends this paradigm by simultaneously searching from the start and goal states to reduce overall search effort. This work expands on both the theoretical foundations and practical capabilities of BiHS, aiming to clarify when bidirectional search is effective and how it can be made scalable in practice. It presents a unifying theoretical perspective on BiHS with front-to-end (F2E) heuristics and shows empirically that modern algorithms often operate close to known theoretical lower bounds on the number of node expansions. It further analyzes front-to-front (F2F) heuristics, highlighting the tradeoff between their ability to reduce node expansions and their additional computational overhead. Beyond theoretical insights, this work addresses scalability through parallelization and external-memory search, introducing a general framework that enables different BiHS algorithms to efficiently exploit multicore architectures and external storage, allowing them to solve substantially larger problem instances. Finally, it extends BiHS to the bounded-suboptimal setting, designing algorithms that maintain provable solution quality guarantees while further improving search efficiency. Collectively, these contributions provide a more unified, scalable, and general understanding of bidirectional heuristic search. Speaker: Maxim Bragilovski Supervisor: Prof. Arnon Sturm Date: Monday, 25/5, talk starts at 12:30 Place: Building 37, room 202 Talk title: Automated Support for Domain Model Derivation from User Stories Abstract: Domain models are central artifacts in software development, supporting communication, requirements analysis, and system design. However, deriving domain models from user stories remains a cognitively demanding task, especially for novice modelers, due to ambiguity, incompleteness, and the need to identify relevant classes and associations. This talk presents my research on automated and AI-supported domain model derivation from user stories. I will discuss empirical comparisons between human modelers, rule-based NLP tools, machine learning approaches, and large language models, showing that while automated methods do not fully replace human expertise, LLMs can provide useful initial models and perform competitively in class identification. I will also present recent work on how LLM-generated models affect novice modelers, highlighting both their benefits and the risk of over-reliance, and introduce DOMINO, a multi-agent framework designed to improve the accuracy, consistency, and explainability of generated domain models.