SPEAKER: Debasmita Lohar, Assistant Professor at ITU.
ABSTRACT: Modern cyber-physical systems rely on feedback controllers that must execute efficiently on resource-constrained systems. Numerical precision creates a trade-off between control performance and computational cost: higher precision improves control quality but increases runtime, while lower precision is faster but may degrade performance.
We present the first closed-loop control scheduling framework with per-sample precision switching that determines when to execute a controller in which floating-point precision while keeping the system output within a specified reference band. We formulate this as a multi-objective Mixed-Integer Quadratic Program (MIQP) that accounts for roundoff errors. Experiments on standard control benchmarks show that switching between 32-bit and 16-bit precision reduces runtime by 5.8% on average over always using 32-bit precision and improves control performance by 24.8% over always using 16-bit precision. Our approach also outperforms simple heuristic switching strategies and improves the schedulability of multiple control tasks.
This is a joint work with Debarpita Banerjee and Sumana Ghosh