CNS Core: Small: AccelRITE: Accelerating ReInforcemenT Learning based AI at the Edge Using FPGAs

人工智能(AI)在自动驾驶汽车和机器人等多个领域取得了重大进展。强化学习(RL)是人工智能的一类，它包括使机器能够自学最优决策的算法。然而，RL算法复杂且耗时，不适合需要快速响应的应用。将中央处理单元(CPU)与可配置的集成电路相耦合的异类平台--现场可编程门阵列(现场可编程门阵列)--因其功能而成为实现快速算法的有前途的候选。该项目将针对这些平台开发RL算法的快速实现。该项目的智力优势包括研究和开发创新的优化，这些优化利用了新兴类别的FPGA设备的异构性，并解决了对共享对象的并行访问中的冲突、不规则的内存访问和细粒度加速的开销等挑战。该项目将为关键的人工智能内核开发参数化的性能模型-随机梯度下降(SGD)、共轭梯度、并行哈希表和神经网络，以实现能量性能权衡分析。该项目将开发一种新的基于时空约束图的设计空间探索技术，通过整体考虑RL算法来加速该算法。该项目的更广泛的影响是高效地使用由CPU和FPGA组成的异构体系结构，并结合高速缓存相干存储器来加速用于边缘计算的人工智能。这个项目的成功完成将导致可以部署在现实世界环境中的人工智能应用程序的复杂性显著增加。这将导致自动驾驶汽车、机器人和可穿戴医疗设备等支持人工智能的设备的能力大幅提高。该项目还将构成适合纳入研究生和本科生课程的材料。在该项目中开发的所有软件将发布在GitHub上：https://github.com/pgroupATusc.软件版本将在授权期结束后保持不少于3年的时间。这一奖励反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial Intelligence (AI) has led to significant progress in several domains such as self-driving cars and robotics. Reinforcement Learning (RL) is a class of AI that includes algorithms that enable machines to teach themselves optimal decision making. However, RL algorithms are complex and time-consuming, which render them unsuitable for applications that require fast response. Heterogeneous platforms, which couple a Central Processing Unit (CPU) with an integrated circuit that can be configured - Field Programmable Gate Arrays (FPGA) are promising candidates for implementing fast algorithms due to their capabilities. The project will develop fast implementations of RL algorithms targeting such platforms. The intellectual merits of the project include the research and development of innovative optimizations that exploit the heterogeneity of the emerging class of FPGA devices and address challenges such as conflicts in parallel accesses to shared objects, irregular memory accesses, and overheads in fine grained acceleration. The project will develop parameterized performance models for key AI kernels – Stochastic Gradient Descent (SGD), conjugate gradient, parallel hash tables, and neural networks, to enable energy-performance trade-off analysis. The proposed project will develop a novel spatiotemporal constraint graph-based design space exploration technique to accelerate RL algorithms by taking a holistic view of the algorithm.The broader impact of the project is in efficient use of heterogeneous architectures consisting of CPUs and FPGAs coupled with cache coherent memory for accelerating AI for edge computing. Successful completion of this project will lead to significant increase in the complexity of AI applications that can be deployed in real world environments. This will lead to a dramatic improvement in the capabilities of AI enabled devices such as self-driving cars, robotics, and wearable healthcare devices. The project will also constitute materials appropriate for inclusion in graduate and undergraduate courses.All software developed in the project will be posted on github at: https://github.com/pgroupATusc. Software releases will be maintained for a period of not less than 3 years after the conclusion of the grant.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.