Processing Systems Using Markov Decision Processes 1 Georgia Institute of Technology USA 2 University of Maryland USA 3 Tampere University of Technology Finland With contributions from ID: 624545
Download Presentation The PPT/PDF document "Design and Implementation of Adaptive Si..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Design and Implementation of Adaptive SignalProcessing Systems Using Markov DecisionProcesses
1. Georgia Institute of Technology, USA2. University of Maryland, USA3. Tampere University of Technology, Finland
With contributions from: Lin Li2, Adrian E. Sapio2, Jiahao Wu2, Yanzhou Liu2, Kyunghun Lee2, and Shuvra S. Bhattacharyya2,3
Marilyn
Wolf
1Slide2
Outline
2Motivation and ContributionBackgroundFramework for Design and Implementation of Adaptive Signal Processing SystemsCase Study of Channelizer/receiver ApplicationConclusionsSlide3
Motivation and Contribution
3Modern signal processing applications impose increasing demands of:AdaptivityEfficiencyReconfigurabilityFlexibilityChallenges at many levels of system design, implementation and optimization:Dynamically-changing working scenariosStringent constraints on energy-efficiency and real-time performanceMultidimensional design spaceSlide4
Motivation and Contribution
4This research aims at developing a framework for design and implementation of adaptive embedded signal processing systems that integratesautomated, MDP-based generation of optimal reconfiguration policiesdataflow-based application modelingimplementation of embedded control software that carries out the generated reconfiguration policiesSlide5
5
Background Markov Decision Processes (MDPs)
Probabilistic transitions combined with inputs.Given an input at a state, next state is chosen probabilistically.A policy p defines the actions in each state s.Optimal policy maximizes rewards.Slide6
6
Basic concepts of dataflow modeling:Digital Signal Processing (DSP) system directed dataflow graphComputational functions nodes (actors)Communication channels between actors edges (FIFOs)Actor Firing: Actor execution as a discrete unit of computationToken: The encapsulation of some well-defined amount of data.Consumption/Production Rate: Number of tokens consumed/produced from/to the input/output FIFO during one actor firing.
Background
Dataflow
ModelingSlide7
BackgroundLightWeight DataFlow (LWDF)
7LightWeight DataFlow (LWDF): a programming methodology for integration, experimentation, and optimization with dataflow modeling approaches.Actor Mode: Determines the dataflow behavior of the actor.Enable function: Checks actor firing condition according to its current mode. This function can be bypassed at run time if static scheduling analysis can ensure the result.
Invoke function: Executes an actor firing according to its current mode.Lightweight Dataflow Environment (LIDE): Provides a compact set of application programming interfaces (APIs) that is used for constructing, connecting, and executing dataflow components such as actors, edges, and graphs.LIDE APIs have been implemented in a variety of implementation languages, including C, Verilog, and CUDASlide8
Framework for Adaptive Signal Processing Systems
8We propose a novel framework, called
Hierarchical MDP framework for Compact System-level Modeling (HMCSM).Slide9
Framework for Adaptive Signal Processing Systems
9Hierarchical MDP SubsystemA single Markov decision process (MDP) is transformed into a hierarchy of multiple MDPs that can be independently solved.Such decomposition into a collection of simpler MDPs leads to more efficient design optimization.The stochastic models of the environment and system include, for each of the environment and system, the definition of the state space and the state transition matrix (STM). Control Actions represent the set of possible reconfiguration operations.
The Reward Function maps state-action pairs into scores that assess the utility of performing the associated control action during the given state.Slide10
Framework for Adaptive Signal Processing Systems
10Parameterized LIDE ImplementationDataflow graph implementation of the application developed using LIDEParameter updates are made by setting (at design time or run time) appropriate variables in this implementationConfiguration Control Machine (CCM)Determines, based on the current environmental state and system state, whether or not to perform a dynamic reconfiguration operationDetermines the specific reconfiguration operation that is to be applied to the system if reconfiguration is to be performed
Policy Mapping Engine (MPE)Translates control actions into updates to dynamic parameters in the embedded softwareSlide11
Case Study of Channelizer/receiver Application (1)
11Slide12
Case Study of Channelizer/receiver Application (2)
12 PSDF specification of channelizer subsystem.
Init Graph: modeling construct in parameterized synchronous dataflow (PSDF) for reconfiguration functionality (determinationand propagation of new parameter values).Body Graph: modeling construct in PSDF for core signal processing functionality associated with an application.Slide13
Case Study of Channelizer/receiver Application (3)
13
Platform for Channelizer: Raspberry Pi 3 Model BDevice for Power Measurement: Tektronix Keithley Series 2280 Precision Measurement DC Power Supply Average processing power of all the available configurations.Slide14
Case Study of Channelizer/receiver Application (4)
14Simulation results for MDP-I.
Comparison among MDP-generated policies and fixed-configuration designs.In addition to providing better energy efficiency compared to the fixed configuration designs, our MDP approach can be configured systematically to generate a much larger set of trade-off options (Pareto-optimized fronts)
ensures optimality (with respect to the given reward function)Slide15
15
Solver running timeMDP Solver: MATLAB-based open source solver, named MDPSOLVEPlatform for MDP Solver: Processor: i7-4710HQ running at 2.50GHzRAM: 12.0 GBOS: 64-bit Windows 10MDP-I solver running time: 294msMDP-II-a solver running time: 50.8msMDP-II-b solver running time: 41.5msIn a deployment with a fixed processing system (MDP-II-b) and changing external environment (MDP-II-a), the hierarchical MDP scheme reduces the solver time from 294ms to 50.8ms, which is a factor of over 5.7X smaller.
Model sizeMDP-I: 1.63MBMDP-II: 265kBMDP-II reduces model size by a factor of over 6.1X.Slide16
Conclusions
16We propose the HMCSM framework for design and implementation of adaptive embedded signal processing systems.HMCSM stands for Hierarchical MDP framework for Compact System-level Modeling.HMCSM provides a structured methodology that integrates dataflow
methods; MDP formulation using compact, hierarchical models; optimal policy generation at design time; and dynamic, system-level reconfiguration at run time.We demonstrate the effectiveness of our new MDP-based system design framework through experiments with an adaptive wireless communications receiver.Slide17
Thanks!
17