Learning to Control Binary Decision Diagrams with Reinforcement Learning
From International Center for Computational Logic
Learning to Control Binary Decision Diagrams with Reinforcement Learning
Talk by Luke Slater
- Location: APB 3027
- Start: 5. November 2026 at 11:00 am
- End: 5. November 2026 at 12:00 pm
- Research group: Knowledge-Based Systems
- Event series: Research Seminar Logic and AI
- iCal
Binary Decision Diagrams (BDDs) are a powerful representation of Boolean functions, supporting efficient logical operations and reasoning over large state spaces. However, their practical effectiveness depends heavily on the ordering of variables, with poor orderings potentially causing an exponential increase in representation size. A major limitation of standard BDD-based computation is its reliance on computationally expensive dynamic variable reordering heuristics, which perform extensive search to optimise the size of a static set of diagrams. These methods are fundamentally reactive, aiming to repair variable orderings once the diagrams have already grown substantially large, rather than preventing this growth in the first place. In this talk, I will present my ongoing PhD research on developing a reinforcement learning controller that continuously regulates variable ordering throughout BDD construction and manipulation. Rather than relying on expensive search to repair large diagrams, the objective is to learn a predictive reordering policy that makes decisions directly from the structure of the current diagrams, proactively maintaining compact representations as computation progresses.
