In this paper, we study reinforcement learning (RL) algorithms to solve real-world decision problems with the objective of maximizing the long-term reward as well as satisfying cumulative constraints. We propose a novel first-order policy optimization method, Interior-point Policy Optimization (IPO), which augments the objective with logarithmic barrier functions, inspired by the interior-point method. Our proposed method is easy to implement with performance guarantees and can handle general types of cumulative multiconstraint settings. We conduct extensive evaluations to compare our approach with state-of-the-art baselines. Our algorithm outperforms the baseline algorithms, in terms of reward maximization and constraint satisfaction.
more »
« less
DTRL: Decision Tree-based Multi-Objective Reinforcement Learning for Runtime Task Scheduling in Domain-Specific System-on-Chips
Domain-specific systems-on-chip (DSSoCs) combine general-purpose processors and specialized hardware accelerators to improve performance and energy efficiency for a specific domain. The optimal allocation of tasks to processing elements (PEs) with minimal runtime overheads is crucial to achieving this potential. However, this problem remains challenging as prior approaches suffer from non-optimal scheduling decisions or significant runtime overheads. Moreover, existing techniques focus on a single optimization objective, such as maximizing performance. This work proposes DTRL, a decision-tree-based multi-objective reinforcement learning technique for runtime task scheduling in DSSoCs. DTRL trains a single global differentiable decision tree (DDT) policy that covers the entire objective space quantified by a preference vector. Our extensive experimental evaluations using our novel reinforcement learning environment demonstrate that DTRL captures the trade-off between execution time and power consumption, thereby generating a Pareto set of solutions using a single policy. Furthermore, comparison with state-of-the-art heuristic–, optimization–, and machine learning-based schedulers shows that DTRL achieves up to 9× higher performance and up to 3.08× reduction in energy consumption. The trained DDT policy achieves 120 ns inference latency on Xilinx Zynq ZCU102 FPGA at 1.2 GHz, resulting in negligible runtime overheads. Evaluation on the same hardware shows that DTRL achieves up to 16% higher performance than a state-of-the-art heuristic scheduler.
more »
« less
- Award ID(s):
- 2114499
- PAR ID:
- 10537760
- Publisher / Repository:
- ACM Digital Library
- Date Published:
- Journal Name:
- ACM Transactions on Embedded Computing Systems
- Volume:
- 22
- Issue:
- 5s
- ISSN:
- 1539-9087
- Page Range / eLocation ID:
- 1 to 22
- Subject(s) / Keyword(s):
- Computer systems organization → System on a chip • Hardware → On-chip re- source management Domain-specific system-on-chip, task scheduling, reinforcement learning, decision trees, resource management, multi-objective optimization
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We propose a novel policy gradient method for multi-agent reinforcement learning, which leverages two different variance-reduction techniques and does not require large batches over iterations. Specifically, we propose a momentum-based decentralized policy gradient tracking (MDPGT) where a new momentum-based variance reduction technique is used to approximate the local policy gradient surrogate with importance sampling, and an intermediate parameter is adopted to track two consecutive policy gradient surrogates. MDPGT provably achieves the best available sample complexity of O(N -1 e -3) for converging to an e-stationary point of the global average of N local performance functions (possibly nonconcave). This outperforms the state-of-the-art sample complexity in decentralized model-free reinforcement learning and when initialized with a single trajectory, the sample complexity matches those obtained by the existing decentralized policy gradient methods. We further validate the theoretical claim for the Gaussian policy function. When the required error tolerance e is small enough, MDPGT leads to a linear speed up, which has been previously established in decentralized stochastic optimization, but not for reinforcement learning. Lastly, we provide empirical results on a multi-agent reinforcement learning benchmark environment to support our theoretical findings.more » « less
-
As the era of high frequency, single core processors have come to a close, the new paradigm of many core processors has come to dominate. In response to these systems, asynchronous multitasking runtime systems have been developed as a promising solution to efficiently utilize these newly available hardware. Asynchronous multitasking runtime systems work by dividing a problem into a large number of fine grained tasks. However, as the number of tasks created increase, the overheads associated with task creation and management cannot be ignored. Task inlining, a method where the parent thread consumes a child thread, enables the runtime system to achieve the balance between parallelism and its overhead. As largely impacted by different processor architectures, the decision of task inlining is dynamic in nature. In this research, we present adaptive techniques for deciding, at runtime, whether a particular task should be inlined or not. We present two policies, a baseline policy that makes inlining decision based on a fixed threshold and an adaptive policy which decides the threshold dynamically at runtime. We also evaluate and justify the performance of these policies on different processor architectures. To the best of our knowledge, this is the first study of the impacts of adaptive policy at runtime for task inlining in an asynchronous multitasking runtime system on different processor architectures. From experimentation, we find that the baseline policy improves the execution time from 7.61% to 54.09%. Furthermore, the adaptive policy improves over the baseline policy by up to 74%.more » « less
-
We propose a systematic application-specific hardware design methodology for designing Spiking Neural Network (SNN), SNNOpt, which consists of three novel phases: 1) an Olliver-Ricci-Curvature (ORC)-based architecture-aware network partitioning, 2) a reinforcement learning mapping strategy, and 3) a Bayesian optimization algorithm for NoC design space exploration. Experimental results show that SNNOpt achieves a 47.45% less runtime and 58.64% energy savings over state-of-the-art approaches.more » « less
-
null (Ed.)Internet of Things (IoT) sensors often operate in unknown dynamic environments comprising latency-sensitive data sources, dynamic processing loads, and communication channels of unknown statistics. Such settings represent a natural application domain of reinforcement learning (RL), which enables computing and learning decision policies online, with no a priori knowledge. In our previous work, we introduced a post-decision state (PDS) based RL framework, which considerably accelerates the rate of learning an optimal decision policy. The present paper formulates an efficient hardware architecture for the action evaluation step, which is the most computationally-intensive step in the PDS based learning framework. By leveraging the unique characteristics of PDS learning, we optimize its state value expectation and known cost computational blocks, to speed-up the overall computation. Our experiments show that the optimized circuit is 49 times faster than its software implementation counterpart, and six times faster than a Q-learning hardware accelerator.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3609108
