A synergistic reinforcement learning-based framework design in driving automation

Abstract

Autonomous driving, which integrates artificial intelligence and the Internet of Things, has piqued the interest of both academics and industry because of its economic and societal benefits. Rigorous accuracy and latency requirements are important for autonomous driving safety. In order to achieve high computation performance in driving automation system, we propose in this paper a heterogeneous multicore AI accelerator (HMAI). At the same time, on the HMAI, how to allocate a large number of real-time tasks to different accelerators remains a notable problem that is worth considering. Theoretically, this problem is NP-complete, and always solved using heuristic-based and guided random-search-based algorithms. However, the global state of HMAI cannot be considered comprehensively in these algorithms, which usually leads to suboptimal allocations. In this paper, we propose FlexAI, a predictive and global scheduling mechanism on HMAI. Specifically, the proposed scheduling algorithm that is based upon deep reinforcement learning (RL). In order to evaluate the quality of strategies produced by RL agent and update the observation of the scheduling agent, two scheduling metrics are proposed: Global State Value (Gvalue), Matching Score (MS) which pays attention to the requirements of various tasks in driving automation system like emergency level. In the experimental, FlexAI achieves up to 80% execution time reduction and 99% resource utilization improvement compared with Min-min, ATA in heuristics, and genetic algorithms, simulated annealing in guided random-search-based algorithms, and unscheduled case.