DOCSLIB.ORG

Classical Control

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Classical Control Classical Control Topics covered: Modeling. ODEs. Linearization. Laplace transform. Transfer functions. Block diagrams. Mason’s Rule. Time response specifications. Effects of zeros and poles. Stability via Routh-Hurwitz. Feedback: Disturbance rejection, Sensitivity, Steady-state tracking. PID controllers and Ziegler-Nichols tuning procedure. Actuator saturation and integrator wind-up. Root locus. Frequency response--Bode and Nyquist diagrams. Stability Margins. Design of dynamic compensators. Classical Control – Prof. Eugenio Schuster – Lehigh University 1 Classical Control Text: Feedback Control of Dynamic Systems, 4th Edition, G.F. Franklin, J.D. Powel and A. Emami-Naeini Prentice Hall 2002. Classical Control – Prof. Eugenio Schuster – Lehigh University 2 1 What is control? For any analysis we need a mathematical MODEL of the system Model → Relation between gas pedal and speed: 10 mph change in speed per each degree rotation of gas pedal Disturbance → Slope of road: 5 mph change in speed per each degree change of slope Block diagram for the cruise control plant: w Slope (degrees) 0.5 y =10(u − 0.5w) Control - Output speed (degrees) (mph) 10 u + y Classical Control – Prof. Eugenio Schuster – Lehigh University 3 What is control? Open-loop cruise control: w r u = PLANT 10 0.5 yol =10(u − 0.5w) - r 1/10 10 =10( − 0.5w) r u + yol 10 Reference = r − 5w (mph) r = 65,w = 0 ⇒ eol = 0 eol = r − yol = 5w r = 65,w =1⇒ eol = 5mph,eol = 7.69% r − yol w eol [%] = = 500 OK when: r r 1- Plant is known exactly 2- There is no disturbance Classical Control – Prof. Eugenio Schuster – Lehigh University 4 2 What is control? Closed-loop cruise control: w u = 20(r − ycl ) PLANT 0.5 ycl =10(u − 0.5w) 200 5 - + = r − w 10 201 201 1/10 u + r - ycl 1 5 1 ecl = r − ycl = r + w r = 65,w = 0 ⇒ e = % = 0.5% 201 201 cl 201 1 5 5 r − y 1 5 w r = 65,w =1⇒ e = + = 0.69% e [%] = cl = + cl 201 20165 cl r 201 201 r Classical Control – Prof. Eugenio Schuster – Lehigh University 5 What is control? Feedback control can help: • reference following (tracking) • disturbance rejection • changing dynamic behavior LARGE gain is essential but there is a STABILITY limit “The issue of how to get the gain as large as possible to reduce the errors due to disturbances and uncertainties without making the system become unstable is what much of feedback control design is all about” First step in this design process: DYNAMIC MODEL Classical Control – Prof. Eugenio Schuster – Lehigh University 6 3 Dynamic Models MECHANICAL SYSTEMS: F = ma Newton’s law m&x& = u − bx& v = x& velocity a = v& = &x& acceleration b u Vo 1 m Transfer Function v& + v = ⎯⎯→⎯st ⎯⎯⎯st = v=Voe ,u=Uoe d m m Uo s + b m → s dt Classical Control – Prof. Eugenio Schuster – Lehigh University 7 Dynamic Models MECHANICAL SYSTEMS: F = Iα Newton’s law 2 ml θ&& = −lmg sinθ + Tc ω = θ& angular velocity α = ω& = θ&& angular acceleration 2 I = ml moment of inertia g T g T θ&&+ sinθ = c ⎯⎯→⎯ θ&&+ θ = c Linearization l ml 2 sinθ ≈θ l ml 2 Classical Control – Prof. Eugenio Schuster – Lehigh University 8 4 Dynamic Models g T θ&&+ θ = c l ml 2 Reduce to first order equations: x1 = x2 x1 = θ & g Tc x2 = θ& x = − x + &2 l 1 ml 2 ⎡x ⎤ T ⎡ 0 1⎤ ⎡0⎤ x ≡ 1 ,u ≡ c ⇒ x = ⎢ g ⎥x + u State Variable ⎢x ⎥ ml 2 & − 0 ⎢1⎥ Representation ⎣ 2 ⎦ ⎣⎢ l ⎦⎥ ⎣ ⎦ General case: x& = Fx + Gu y = Hx + Ju Classical Control – Prof. Eugenio Schuster – Lehigh University 9 Dynamic Models ELECTRICAL SYSTEMS: Kirchoff’s Current Law (KCL): The algebraic sum of currents entering a node is zero at every instant Kirchoff’s Voltage Law (KVL) The algebraic sum of voltages around a loop is zero at every instant Resistors: vR (t) = RiR (t) ⇔ iR (t) = GvR (t) i i R iC L CitCapacitors: + + + t v v dv (t) 1 R vC L i (t) = C C ⇔ v (t) = i (τ )dτ + v (0) C C ∫ C C dt C 0 Inductors: di (t) 1 t v (t) = L L ⇔ i (t) = v (τ )dτ + i (0) L L ∫ L L dt L 0 Classical Control – Prof. Eugenio Schuster – Lehigh University 10 5 Dynamic Models ELECTRICAL SYSTEMS: OP AMP: vO = A(vp − vn ), A → ∞ i + p + vp RO iO v vp = vn RI + O + in + i = i = 0 v A(vp-vn) p n n - - To work in the linear mode we need FEEDBACK!!! Classical Control – Prof. Eugenio Schuster – Lehigh University 11 Dynamic Models ELECTRICAL SYSTEMS: KCL: R2 R 1 C dvO 1 1 v vO = vO − vI 1 dt R C R C - 2 1 + 1 t R = ∞ (OC) ⇒ v ()t = v (0 )− v (μ)dμ 2 O O ∫0 I R1C v v 1 1 K ∫ O K = − RC Inverting integrator Classical Control – Prof. Eugenio Schuster – Lehigh University 12 6 Dynamic Models ELECTRO-MECHANICAL SYSTEMS: DC Motor torque armature current T = Ktia e = Keθ&m emf shaft velocity J mθ&&m = −bθ&m + T di − v + R i + L a + e = 0 a a a dt Obtain the State Variable Representation Classical Control – Prof. Eugenio Schuster – Lehigh University 13 Dynamic Models HEAT-FLOW: Heat Flow Temperature Difference 1 q = ()T −T R 1 2 1 T& = q C Thermal capacitance Thermal resistance 1 ⎛ 1 1 ⎞ T&I = ⎜ + ⎟()To −TI CI ⎝ R1 R2 ⎠ Classical Control – Prof. Eugenio Schuster – Lehigh University 14 7 Dynamic Models FLUID-FLOW: Mass rate Mass Conservation law m& = win − wout Inlet mass flow Outlet mass flow 1 m = ρAh& ⇒ h& = (w − w ) & ρA in out A: area of the tank ρ: density of fluid h: height of water Classical Control – Prof. Eugenio Schuster – Lehigh University 15 Linearization Dynamic System: x& = f (x,u) 0 = f (xo ,uo ) Equilibrium Denote δx = x − xo ,δu = u − uo δx& = f ()xo + δx,uo + δu Taylor Expansion ∂f ∂f δx ≈ f ()x ,u + δx + δu & o o ∂x ∂u xo ,uo xo ,uo ∂f ∂f F ≡ ,G ≡ ⇒ δx ≈ Fδx + Gδu ∂x ∂u & xo ,uo xo ,uo Classical Control – Prof. Eugenio Schuster – Lehigh University 16 8 Linearization δx& ≈ Fδx + Gδu ⎡∂f1 ∂f1 ⎤ ⎡ ∂f1 ∂f1 ⎤ ⎢∂x L ∂x ⎥ ⎢∂u L ∂u ⎥ ∂f ⎢ 1 n ⎥ ∂f ⎢ 1 m ⎥ F ≡ = ⎢ M M ⎥ ,G ≡ = ⎢ M M ⎥ ∂x x ,u ∂u x ,u o o ⎢∂fn ∂fn ⎥ o o ⎢∂fn ∂fn ⎥ ⎢ L ⎥ ⎢ L ⎥ ∂x1 ∂xn ∂u1 ∂um ⎣ ⎦ xo ,uo ⎣ ⎦ xo ,uo k g Example: Pendulum with friction θ&&+ θ& + sinθ = 0 m l ⎡ 0 1 ⎤ x = g k x & ⎢− cos x − ⎥ ⎢ 1 ⎥ ⎣ l m⎦ xo Classical Control – Prof. Eugenio Schuster – Lehigh University 17 Laplace Transform • Function f(t) of time bt – Piecewise continuous and exponential order f (t) < Ke ∞ α + j∞ −st −1 1 st F(s) = ∫ f (t)e dt L [F(s)]= f (t) = ∫ F(s)e ds π 0- 2 j α − j∞ – 0- limit is used to capture transients and discontinuities at t=0 – s is a complex variable (σ+jω) • There is a need to worry about regions of convergence of the integral – Units of s are sec-1=Hz • A frequency –If f(t) is volts (amps) then F(s) is volt-seconds (amp-seconds) Classical Control – Prof. Eugenio Schuster – Lehigh University 18 9 Laplace transform examples • Step function – unit Heavyside Function ⎧0, for t < 0 • After Oliver Heavyside (1850-1925) u(t) = ⎨ ⎩1, for t ≥ 0 ∞ ∞ ∞ ∞ e−st e−(σ + jω)t 1 F(s) = ∫u(t)e−stdt = ∫ e−stdt = − = − = if σ > 0 s σ + jω s 0− 0− 0 0 • Exponential function • After Oliver Exponential (1176 BC- 1066 BC) ∞ ∞∞ e−(s+α )t 1 F(s) = ∫∫e−αte−stdt = e−(s+α )tdt = − = if σ > α s +α s +α 0 0 0 • Delta (impulse) function δ(t) ∞ F(s) = ∫δ (t)e−stdt =1 for all s 0− Classical Control – Prof. Eugenio Schuster – Lehigh University 19 Laplace Transform Pair Tables Signal Waveform Transform impulse δ (t) 1 step u(t) 1 s 1 ramp tu(t) s2 1 exponential e−αtu(t) s+α t 1 damped ramp te−α u(t) (s+α )2 β sine sin(βt)u(t) s2+β 2 s cosine cos(βt)u(t) s2+β 2 damped sine t β e−α sin()βt u(t) (s+α )2+β 2 damped cosine t s+α e−α cos()βt u(t) (s+α )2+β 2 Classical Control – Prof. Eugenio Schuster – Lehigh University 20 10 Laplace Transform Properties Linearity: (absolutely critical property) L{}Af1(t) + Bf2(t) = AL{f1 (t)}+ BL{f2 (t)}= AF1(s) + BF2(s) ⎧t ⎫ F(s) Integration property: L⎨∫ f (τ )dτ ⎬ = ⎩0 ⎭ s ⎧df (t)⎫ Differentiation property: L⎨ ⎬ = sF(s) − f (0−) ⎩ dt ⎭ ⎧ 2 ⎫ ⎪d f (t)⎪ 2 L = s F(s) − sf (0−) − f ′(0−) ⎨ 2 ⎬ ⎩⎪ dt ⎭⎪ ⎪⎧d m f (t)⎪⎫ = m F(s) − sm−1 f (0−) − sm−2 f ′(0−) − − f (m) (0−) L⎨ m ⎬ s L ⎩⎪ dt ⎭⎪ Classical Control – Prof. Eugenio Schuster – Lehigh University 21 Laplace Transform Properties Translation properties: s-domain translation: L{e−αt f (t)} = F(s +α) −as t-domain translation: L{f (t − a)u(t − a)}= e F(s) for a > 0 Initial Value Property: lim f (t) = lim sF(s) t→0+ s→∞ Fina l Va lue Proper ty: lim f (t) = lim sF(s) t→∞ s→0 If all poles of F(s) are in the LHP Classical Control – Prof. Eugenio Schuster – Lehigh University 22 11 Laplace Transform Properties 1 s Time Scaling: L{ f (at)} = F( ) a a dF(s) Multiplication by time: L{tf (t)} = − ds t Convolution: L{ f (τ )g(t −τ )dτ} = F(s)G(s) ∫0 1 σ + jω Time product: L{ f (t)g(t)} = F(s)G(s − λ)dλ 2πj ∫σ − jω Classical Control – Prof.
Load more

Recommended publications
  • Control Theory
    Control theory S. Simrock DESY, Hamburg, Germany Abstract In engineering and mathematics, control theory deals with the behaviour of dynamical systems. The desired output of a system is called the reference. When one or more output variables of a system need to follow a certain ref- erence over time, a controller manipulates the inputs to a system to obtain the desired effect on the output of the system. Rapid advances in digital system technology have radically altered the control design options. It has become routinely practicable to design very complicated digital controllers and to carry out the extensive calculations required for their design. These advances in im- plementation and design capability can be obtained at low cost because of the widespread availability of inexpensive and powerful digital processing plat- forms and high-speed analog IO devices. 1 Introduction The emphasis of this tutorial on control theory is on the design of digital controls to achieve good dy- namic response and small errors while using signals that are sampled in time and quantized in amplitude. Both transform (classical control) and state-space (modern control) methods are described and applied to illustrative examples. The transform methods emphasized are the root-locus method of Evans and fre- quency response. The state-space methods developed are the technique of pole assignment augmented by an estimator (observer) and optimal quadratic-loss control. The optimal control problems use the steady-state constant gain solution. Other topics covered are system identification and non-linear control. System identification is a general term to describe mathematical tools and algorithms that build dynamical models from measured data.
    [Show full text]
  • Graphical Control Tools to Design Control Systems
    Advances in Engineering Research, volume 141 5th International Conference on Mechatronics, Materials, Chemistry and Computer Engineering (ICMMCCE 2017) Graphical Control Tools to Design Control Systems Huanan Liu1,a, , Shujian Zhao1,b, and Dongmin Yu 1,c,* 1 Department of Electrical Engineering, Northeast Electric Power University, Jilin 132012, China a [email protected], b [email protected], c [email protected], *corresponding author: Dongmin Yu email:[email protected] Keywords: Root Locus; Frequency Tools; Bode Diagram; Nyquist Diagram Abstract: Engineers are sometimes asked to analyze the feasibility of existing control systems or design a new system. The conventional analysis and design methods heavily depend on mathematical calculation. In order to minimize calculation, graphical control is proposed. This paper aims to introduce three different graphical control tools (Root Locus, Bode diagram and Nyquist diagram) and then use these tools to analyze and design the control system. Finally, some realistic case studies are used to compare the advantages and disadvantages of graphical tools. The innovation of this paper is adding hand-drawn graphical control to greatly reduce the calculation complexity 1. List of Symbols -1 Ts-settling time (s) ߞ-damping ratio ߱n-natural frequency (rad.s ) %OS-percentage of overshoot Ess-steady state error K-gain of the system Z1, Z2-zeros of the system P1, P2-poles of the system Tp-peak time (s) M-magnitude of the phasor Φ-phase angle of phasor (degrees) Θ-angle of asymptote Z-the number of closed loop poles P-the number of open- loop poles N- the number of counter clockwise rotation PM-phase margin (degrees) GM-gain margin h1, h2- feedback coefficient 2.
    [Show full text]
  • Discrete Time Control Systems
    Discrete Time Control Systems Lino Guzzella Spring 2013 0-0 1 Lecture — Introduction Inherently Discrete-Time Systems, example bank account Bank account, interest rates r+ > 0 for positive, r− > 0 for negative balances (1 + r+)x(k)+ u(k), x(k) > 0 x(k +1) = (1) (1 + r−)x(k)+ u(k), x(k) < 0 where x(k) is the account’s balance at time k and u(k) is the ∈ ℜ ∈ ℜ amount of money that is deposited to (u(k) > 0) or withdrawn from (u(k) < 0) the account. 1 In general such systems are described by a difference equation of the form x(k +1)= f(x(k),u(k)), x(k) n, u(k) m, f : n×m n ∈ ℜ ∈ ℜ ℜ → ℜ (2) with which an output equation of the form y(k)= g(x(k),u(k)), y(k) p, g : n×m p (3) ∈ ℜ ℜ → ℜ is often associated. You will learn how continuous-time systems can be transformed to a form similar to that form. Note that there is a fundamental difference between inherently discrete time systems and such approximations: for the former there is no meaningful interpretation of the system behavior in between the discrete time instances k = 1, 2,... , while the latter have a clearly { } defined behavior also between two “sampling” events. 2 Discrete-Time Control Systems Most important case: continuous-time systems controlled by a digital computer with interfaces (“Discrete-Time Control” and “Digital Control” synonyms). Such a discrete-time control system consists of four major parts: 1 The Plant which is a continuous-time dynamic system.
    [Show full text]
  • Root Locus Analysis & Design
    Root Locus Analysis & Design • A designer would like: – To know if the system is absolutely stable and the degree of stability. – To predict a system’s performance by an analysis that does NOT require the actual solution of the differential equations. – The analysis to indicate readily the manner or method by which this system must be adjusted or compensated to produce the desired performance characteristics. Mechatronics K. Craig Root Locus Analysis and Design 1 • Two Methods are available: – Root-Locus Approach – Frequency-Response Approach • Root Locus Approach – Basic characteristic of the transient response of a closed-loop system is closely related to the location of the closed-loop poles. – If the system has a variable loop gain, then the location of the closed-loop poles depends on the value of the loop gain chosen. Mechatronics K. Craig Root Locus Analysis and Design 2 – It is important to know how the closed-loop poles move in the s plane as the loop gain is varied. – From a Design Viewpoint: • Simple gain adjustment may move the closed-loop poles to desired locations. The design problem then becomes the selection of an appropriate gain value. • If gain adjustment alone does not yield a desired result, addition of a compensator to the system is necessary. – The closed-loop poles are the roots of the closed-loop system characteristic equation. Mechatronics K. Craig Root Locus Analysis and Design 3 – The Root Locus Plot is a plot of the roots of the characteristic equation of the closed-loop system for all values of a system parameter, usually the gain; however, any other variable of the open- loop transfer function may be used.
    [Show full text]
  • Nyquist Stability Criteria
    NPTEL >> Mechanical Engineering >> Modeling and Control of Dynamic electro-Mechanical System Module 2- Lecture 14 Nyquist Stability Criteria Dr. Bis ha kh Bhatt ac harya Professor, Department of Mechanical Engineering IIT Kanpur Joint Initiative of IITs and IISc - Funded by MHRD NPTEL >> Mechanical Engineering >> Modeling and Control of Dynamic electro-Mechanical System Module 2- Lecture 14 This Lecture Contains Introduction to Geometric Technique for Stability Analysis Frequency response of two second order systems Nyquist Criteria Gain and Phase Margin of a system Joint Initiative of IITs and IISc - Funded by MHRD NPTEL >> Mechanical Engineering >> Modeling and Control of Dynamic electro-Mechanical System Module 2- Lecture 14 Introduction In the last two lectures we have considered the evaluation of stability by mathema tica l eval uati on of th e ch aract eri sti c equati on. Rth’Routh’s tttest an d Kharitonov’s polynomials are used for this purpose. There are several geometric procedures to find out the stability of a system. These are based on: Nyquist Plot Root Locus Plot and Bode plot The advantage of these geometric techniques is that they not only help in checking the stability of a system, they also help in designing controller for the systems. Joint Initiative of IITs and IISc ‐ Funded by 3 MHRD NPTEL >> Mechanical Engineering >> Modeling and Control of Dynamic electro-Mechanical System Module 2- Lecture 14 NitNyquist Plo t is bdbased on Frequency Response of a TfTransfer FtiFunction. CidConsider two transfer functions as follows: s 5 s 5 T (s) ; T (s) 1 s2 3s 2 2 s2 s 2 The two functions have identical zero.
    [Show full text]
  • Root Locus Techniques 8
    E1C08 11/02/2010 10:23:11 Page 387 Root Locus Techniques 8 Chapter Learning Outcomes After completing this chapter the studentApago will be ablePDF to: Enhancer Define a root locus (Sections 8.1–8.2) State the properties of a root locus (Section 8.3) Sketch a root locus (Section 8.4) Find the coordinates of points on the root locus and their associated gains (Sections 8.5–8.6) Use the root locus to design a parameter value to meet a transient response specification for systems of order 2 and higher (Sections 8.7–8.8) Sketch the root locus for positive-feedback systems (Section 8.9) Find the root sensitivity for points along the root locus (Section 8.10) Case Study Learning Outcomes You will be able to demonstrate your knowledge of the chapter objectives with case studies as follows: Given the antenna azimuth position control system shown on the front endpapers, you will be able to find the preamplifier gain to meet a transient response specification. Given the pitch or heading control system for the Unmanned Free-Swimming Submersible vehicle shown on the back endpapers, you will be able to plot the root locus and design the gain to meet a transient response specification. You will then be able to evaluate other performance characteristics. 387 E1C08 11/02/2010 10:23:12 Page 388 388 Chapter 8 Root Locus Techniques 8.1 Introduction Root locus, a graphical presentation of the closed-loop poles as a system parameter is varied, is a powerful method of analysis and design for stability and transient response (Evans, 1948; 1950).
    [Show full text]
  • Bicycle and Motorcycle Dynamics - Wikipedia, the Free Encyclopedia 16/1/22 上午 9:00
    Bicycle and motorcycle dynamics - Wikipedia, the free encyclopedia 16/1/22 上午 9:00 Bicycle and motorcycle dynamics From Wikipedia, the free encyclopedia Bicycle and motorcycle dynamics is the science of the motion of bicycles and motorcycles and their components, due to the forces acting on them. Dynamics is a branch of classical mechanics, which in turn is a branch of physics. Bike motions of interest include balancing, steering, braking, accelerating, suspension activation, and vibration. The study of these motions began in the late 19th century and continues today.[1][2][3] Bicycles and motorcycles are both single-track vehicles and so their motions have many fundamental attributes in common and are fundamentally different from and more difficult to study than other wheeled vehicles such as dicycles, tricycles, and quadracycles.[4] As with unicycles, bikes lack lateral stability when stationary, and under most circumstances can only remain upright when moving forward. Experimentation and mathematical analysis have shown that a bike A computer-generated, simplified stays upright when it is steered to keep its center of mass over its model of bike and rider demonstrating wheels. This steering is usually supplied by a rider, or in certain an uncontrolled right turn. circumstances, by the bike itself. Several factors, including geometry, mass distribution, and gyroscopic effect all contribute in varying degrees to this self-stability, but long-standing hypotheses and claims that any single effect, such as gyroscopic or trail, is solely responsible for the stabilizing force have been discredited.[1][5][6][7] While remaining upright may be the primary goal of beginning riders, a bike must lean in order to maintain balance in a turn: the higher the speed or smaller the turn radius, the more lean is required.
    [Show full text]
  • Classical Control Theory
    Classical Control Theory: A Course in the Linear Mathematics of Systems and Control Efthimios Kappos University of Sheffield, School of Mathematics and Statistics January 16, 2002 2 Contents Preface 7 1 Introduction: Systems and Control 9 1.1 Systems and their control . 9 1.2 The Development of Control Theory . 12 1.3 The Aims of Control . 14 1.4 Exercises . 15 2 Linear Differential Equations and The Laplace Transform 17 2.1 Introduction . 17 2.1.1 First and Second-Order Equations . 17 2.1.2 Higher-Order Equations and Linear System Models . 20 2.2 The Laplace Transform Method . 22 2.2.1 Properties of the Laplace Transform . 24 2.2.2 Solving Linear Differential Equations . 28 2.3 Partial Fractions Expansions . 32 2.3.1 Polynomials and Factorizations . 32 2.3.2 The Partial Fractions Expansion Theorem . 33 2.4 Summary . 38 2.5 Appendix: Table of Laplace Transform Pairs . 39 2.6 Exercises . 40 3 Transfer Functions and Feedback Systems in the Complex s-Domain 43 3.1 Linear Systems as Transfer Functions . 43 3.2 Impulse and Step Responses . 45 3.3 Feedback Control (Closed-Loop) Systems . 47 3.3.1 Constant Feedback . 49 4 CONTENTS 3.3.2 Other Control Configurations . 51 3.4 The Sensitivity Function . 52 3.5 Summary . 57 3.6 Exercises . 58 4 Stability and the Routh-Hurwitz Criterion 59 4.1 Introduction . 59 4.2 The Routh-Hurwitz Table and Stability Criterion . 60 4.3 The General Routh-Hurwitz Criterion . 63 4.3.1 Degenerate or Singular Cases . 64 4.4 Application to the Stability of Feedback Systems .
    [Show full text]
  • Learning Dynamics of Gradient Descent Optimization in Deep Neural Networks Wei WU1*, Xiaoyuan JING1*, Wencai DU2 & Guoliang CHEN3
    SCIENCE CHINA Information Sciences May 2021, Vol. 64 150102:1–150102:15 . RESEARCH PAPER . https://doi.org/10.1007/s11432-020-3163-0 Special Focus on Constraints and Optimization in Artificial Intelligence Learning dynamics of gradient descent optimization in deep neural networks Wei WU1*, Xiaoyuan JING1*, Wencai DU2 & Guoliang CHEN3 1School of Computer Science, Wuhan University, Wuhan 430072, China; 2Institute of Data Science, City University of Macau, Macau 999078, China; 3College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China Received 26 April 2020/Revised 22 August 2020/Accepted 19 November 2020/Published online 8 April 2021 Abstract Stochastic gradient descent (SGD)-based optimizers play a key role in most deep learning models, yet the learning dynamics of the complex model remain obscure. SGD is the basic tool to optimize model parameters, and is improved in many derived forms including SGD momentum and Nesterov accelerated gradient (NAG). However, the learning dynamics of optimizer parameters have seldom been studied. We propose to understand the model dynamics from the perspective of control theory. We use the status transfer function to approximate parameter dynamics for different optimizers as the first- or second-order control system, thus explaining how the parameters theoretically affect the stability and convergence time of deep learning models, and verify our findings by numerical experiments. Keywords learning dynamics, deep neural networks, gradient descent, control model, transfer function Citation Wu W, Jing X Y, Du W C, et al. Learning dynamics of gradient descent optimization in deep neural networks. Sci China Inf Sci, 2021, 64(5): 150102, https://doi.org/10.1007/s11432-020-3163-0 1 Introduction Deep neural networks (DNNs) are well applied to solve recognition problems of complex data including image, voice, text, and video, due to their high-dimensional computing capability.
    [Show full text]
  • Root-Locus Analysis and Design
    Root-Locus Design The root-locus can be used to determine the value of the loop gain K , which results in a satisfactory closed-loop behavior. This is called the proportional compensator or proportional controller and provides gradual response to deviations from the set point. There are practical limits as to how large the gain can be made. In fact, very high gains lead to instabilities. If the root-locus plot is such that the desired performance cannot be achieved by the adjustment of the gain, then it is necessary to reshape the root-loci by adding the additional controllerGc (s)to the open-loop transfer function.Gc (s) must be chosen so that the root-locus will pass through the proper region of the s -plane. In many cases, the speed of response and/or the damping of the uncompensated system must be increased in order to satisfy the specifications. This requires moving the dominant branches of the root locus to the left. The proportional controller has no sense of time, and its action is determined by the present value of the error. An appropriate controller must make corrections based on the past and future values. This can be accomplished by combining proportional with integral action PI or proportional with derivative action PD . One of the most common controllers available commercially is the PID controller. Different processes are suited to different combinations of proportional, integral, and derivative control. The control engineer's task is to adjust the three gain factors to arrive at an acceptable degree of error reduction simultaneously with acceptable dynamic response.
    [Show full text]
  • Design of a Planetary Explorer Third Year Project
    Design of a Planetary Explorer Third Year Project Jack Andrews (ChCh) Aamir Aziz (New) Nicholas Baker (SJC) Alexander Chadwick (SJC) James Coates (BNC) Oliver Cohen (LMH) Liam Donovan (Hert) Hugo Grimmett (BNC) James Hawkes (Keb) Michael May (Worc) Joshua McFarlane (SJC) 2009-2010 Contents 1 Introduction 13 1.1 Mission to Titan . 13 1.2 Titan's Characteristics . 13 1.3 Mission Components . 14 2 Delivery of the Orbiter 15 2.1 Introduction . 15 2.2 Orbital Mechanics . 15 2.2.1 Kepler's Laws . 15 2.2.2 Energy Considerations . 17 2.2.3 Characterising Orbits . 18 2.3 Transfer Orbits . 20 2.4 The N-Body Problem . 20 2.4.1 A Computational Model . 20 2.4.2 Sourcing Planetary Data . 21 2.5 Simulating the Problem in Matlab . 22 2.6 Fly-by Manoeuvres . 22 2.6.1 The Origin of Gravity-Assist Velocity Increase . 22 1 CONTENTS 2.6.2 Calculating the Velocity Increase . 23 2.7 The Oberth Effect . 23 2.8 Time-to-Intercept & Mean Anomaly . 23 2.9 Targetting Manoeuvres . 24 2.10 Planning out the route to Titan . 26 2.11 Launch . 26 2.11.1 Calculating the Relative Launch Velocity . 28 2.11.2 Optimal Launch Window . 29 2.12 The Control System . 31 2.12.1 Implementing the Control System in Matlab . 31 2.13 The Simulation . 32 2.14 Conclusions . 35 3 Attitude Control System 36 3.1 Introduction . 36 3.1.1 Purpose . 36 3.1.2 Control Goal . 37 3.1.3 System Overview . 37 3.1.4 Aim .
    [Show full text]
  • 11.4 the Root Locus Method Hitay O ¨ Zbay Introduction the Root Locus Technique Is Agraphical Tool Used in Feedback Control System Analysis and Design
    11-34 Systems, Controls, Embedded Systems, Energy,and Machines 11.4 The Root Locus Method Hitay O ¨ zbay Introduction The root locus technique is agraphical tool used in feedback control system analysis and design. It has been formally introduced to the engineering communitybyW.R.Evans [3,4], who received the RichardE.Bellman Control Heritage Award from the American Automatic Control Council in 1988 for this major contribution. In ordertodiscuss the root locus method, we must first review the basic definition of bounded input bounded output (BIBO) stabilityofthe standardlinear time invariant feedback system shown in Figure11.22, wherethe plant, and the controller,are represented by their transfer functions P ( s )and C ( s ), respectively1 The plant, P ( s ), includes the physical process to be controlled, as well as the actuator and the sensor dynamics. The feedback system is said to be stable if none of the closed-loop transfer functions, from external inputs r and v to internal signals e and u ,haveany poles in the closed right half plane, C þ : ¼fs 2 C : Reð s Þ > 0 g . Anecessarycondition for feedback system stabilityisthat the closed right half plane zeros of P ( s )(respectively C ( s )) are distinct from the poles of C ( s )(respectively P ( s )). When this condition holds, we saythat there is no unstable pole–zero cancellation in taking the product P ( s ) C ( s ) ¼ : G ( s ), and then checking feedback system stabilitybecomes equivalent to checking whether all the roots of 1 þ G ð s Þ¼0 ð 11: 45Þ are in the open left half plane, C : ¼fs 2 C : Reð s Þ 5 0 g .The roots of Equation (11.1) are the closed-loop system poles.
    [Show full text]