Exploring Machine Learning for Faster Mapping and Scheduling of Automotive Applications on ADAS Platforms

Abstract

In this work, we address the challenge of efficient task mapping and scheduling in Advanced Driver Assistance Systems (ADAS), which are becoming increasingly complex. For this, we explore Mixed-Integer Programming (MIP) combined with Machine Learning (ML) techniques as an alternative to previous work using heuristic algorithms such as simulated annealing or genetic algorithms. Our key contributions include: A simplified MIP formulation of the problem with a novel load-balancing objective. Employing Bayesian optimization to expedite the solving process by finding a better MIP solver configuration, reducing worst-case solving time by 65%. ML-driven decision variable prediction via a Graph Convolutional Network, which is able to fix decision variables with an average precision of up to 77 %, allowing for better branching decisions during the solving process. Experimental results demonstrate the potential for faster solving times, highlighting the value of further integrating machine learning with MIP for advanced ADAS scheduling.