Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Michael Arbib: CS564 - Brain Theory and Artificial Intelligence University of Southern California, Fall 2001 Lecture 19. Systems and Feedback Reading Assignment: TMB2 Section 3.1 and 3.2 Note: Self-study on eigenvalues: TMB2, pp.107,108,111-115. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 1 A System is Defined by Five Elements The set of inputs The set of outputs (These involve a choice by the modeler) The set of states: those internal variables of the system — which may or may not also be output variables — which determine the relationship between input and output. state = the system's "internal residue of the past" The state-transition function : how the state will change when the system is provided with inputs. The output function : what output the system will yield with a given input when in a given state. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 2 From Newton to Dynamic Systems Newton's mechanics describes the behavior of a system on a continuous time scale. Rather than use the present state and input to predict the next state, the present state and input determine the rate at which the state changes. Newton's third law says that the force F applied to the system equals the mass m times the acceleration a. F = ma Position x(t) Velocity v(t) = Acceleration a(t) =x (tx) x(t ) Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 3 Newtonian Systems According to Newton's laws, the state of the system is given by the position and velocity of the particles of the system. We now use u(t) for the input = force; and y(t) (equals x(t)) for the output = position. Note: In general, input, output, and state are more general than in the following, simple example. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 4 State Dynamics With only one particle, the state is the 2-dimensional vector x (t ) q(t) = x ( t ) Then d x(t ) x (t ) and dt d x (t ) dt 1 u (t ) sin ce u (t ) mx m yielding the single vector equation x (t ) 0 1 q (t ) x(t ) 0 0 x(t ) 0 x (t ) 1 u (t ) m The output is given by 1 y(t) = x(t) = 0 x(t ) x (t ) The point of the exercise: Think of the state vector as a single point in a multi-dimensional space. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 5 Linear Systems This is an example of a Linear System: q = A q + B u y = Cq where the state q, input u, and output y are vectors (not necessarily 2dimensional) and A, B, and C are linear operators (i.e., can be represented as matrices). Generally a physical system can be expressed by State Change: q(t) = f(q(t), u(t)) Output: y(t) = g(q(t)) where f and g are general (i.e., possibly nonlinear) functions For a network of leaky integrator neurons: the state mi(t) = arrays of membrane potentials of neurons, the output M(t) = s(mi(t)) = the firing rates of output neurons, obtained by selecting the corresponding membrane potentials and passing them through the appropriate sigmoid functions. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 6 Attractors For all recurrent networks of interest (i.e., neural networks comprised of leaky integrator neurons, and containing loops), given initial state and fixed input, there are just three possibilities for the asymptotic state: 1. The state vector comes to rest, i.e. the unit activations stop changing. This is called a fixed point. For given input data, the region of initial states which settles into a fixed point is called its basin of attraction. 2. The state vector settles into a periodic motion, called a limit cycle. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 7 Strange attractors 3. Strange attractors describe such complex paths through the state space that, although the system is deterministic, a path which approaches the strange attractor gives every appearance of being random. Two copies of the system which initially have nearly identical states will grow more and more dissimilar as time passes. Such a trajectory has become the accepted mathematical model of chaos,and is used to describe a number of physical phenomena such as the onset of turbulence in weather. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 8 Stability The study of stability of an equilibrium is concerned with the issue of whether or not a system will return to the equilibrium in the face of slight disturbances: A is an unstable equilibrium B is a neutral equilibrium C is a stable equilibrium, since small displacements will tend to disappear over time. Note: in a nonlinear system, a large displacement can move the ball from the basin of attraction of one equilibrium to another. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 9 Negative Feedback Controller (Servomechanism) Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 10 Using Spindles to Tell -Neurons if a Muscle Needs to Contract What’s missing in this Scheme? Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 11 Using -Neurons to Set the Resting Length of the Muscle Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 12 Self-study: the elegant extension of this scheme to an agonist-antagonist pair of muscles Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 13 Discrete-Activation Feedforward “Cortex” “Spinal Cord” Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 14 “Ballistic” Correction then Feedback This long latency reflex was noted by Navas and Stark. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 15 The Mass-Spring Model of Muscle Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 16 Hooke’s Law – The Linear Range of Elasticity Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 17 The Mass-Spring Model of Muscle The dynamics of muscle length: mx = mg - u(t) - k(x - xo) - bx mg u(t) -k(x-xo) restoring of law] -bx gravitational force on the mass m controllable force of muscle contraction for a constant k > 0 represents the force of muscle for moderate displacements length x from resting value xo [Hooke's viscosity Use as state the two-dimensional vector x(t) q(t) = . so that x(t) 0 1 x(t) 0 . b . u(t) kxo q(t) = k + g+ m m m m x (t) Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 18 An equilibrium state is one which (for fixed values . of the input) does not change, i.e., for which q (t) = 0. Fix the contractile force u(t) to constant u. .. . x =x =0 so .. . mx = mg - u(t) - k(x - xo) - bx yields 0 = mg - u - k(x - xo) Thus, the equilibrium length xe(u) for given force u satisfies mg - u xe(u) = xo + k (3,4) ... in the domain of linearity of the spring Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 19 Linear Systems . q (t) = Aq(t) + (t) where A is a matrix, and (t) is a general input vector. In a feedback system, (t) will itself be a function of the state q — so will in general depend on q() for times < t. Assume that A is non-singular and thus has inverse A-1. Then an equilibrium state qe for fixed (t) = must satisfy 0 = Aqe + so that A-10 = A-1Aqe + A-1i.e. qe + A-1; and so qe = -A-1,a unique equilibrium point Let p = q - qe be the displacement of the system state from this equilibrium. . . d p = (q - qe) = q = Aq + dt = A(p + qe) + = Ap + Aqe + = Ap Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 20 Eigenvectors and Eigenvalues See TMB2, pp.107,108,111-115 for an exposition of their use in characterizing the stability of a linear system. The key result is: . Theorem. The system q = Aq is stable i.e., q(t) 0 as t for each initial state q(0) if both eigenvalues of A have negative real parts. Moreover, the system will oscillate if the eigenvalues are complex. We now consider how that theory is applied to the mass-spring model. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 21 Stability of the Mass-Spring Model Our mass-spring model 0 0 1 . q (t) = k b q(t) + u(t) kxo - g+ m m m m has equilibrium state xe(u) mg - u qe(u) = where xe(u) = xo + k 0 The eigenvalues are 1 b 1,2 = 2 - + m b 2 m 4k m 1 = 2m -b + b2 - 4km which are real if b2 4km. But since b, k and m are all positive, in this case 0 b2 - 4km < b, and so both 1 and 2 must be negative. Again, if 1 and 2 are complex, their real part is b/2m which is always negative. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 22 Theorem: The mass-spring system with each b, m and k positive is always stable and undergoes damped oscillations just in case b2 < 4km. Large b will prevent the system oscillating, unless it is dominated by a large spring term (large k), for fixed mass m. For b2 < 4km, the oscillations are quickly damped either if b is large (high friction or viscosity) or m is small. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 23 For a given m, and a desired resting length x e, the appropriate force u(xe) must satisfy mg - u(xe) xe = x o + k kxe = mg - u(xe) + kxo so that u(xe) = mg + k(xo-xe) If the mass is m + m then the length achieved will be (m+m)g - u(xe) xo + = xe + m.g/k k Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 24 A simple linear model: the mass-spring model with -style feedback Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 25 Substituting u(xe) = mg + k(xo-xe), we obtain 0 0 1 . q(t) = (k+h) b q(t) + (k+h)xe m m m Thus at equilibrium, (k+h)xe (k+h)x = m m and so our new model has the expected property that the equilibrium length is still xe. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 26 If the mass changes unexpectedly from m to m+m while xe and u(xe) remain fixed, then equation (10) becomes 0 1 x(t) . k b q (t) = x. (t) m+m m+m 0 u(x ) + h(x-x ) kx e e o + g+ m+m m+m However, when we now substitute for u(xe) we must note that this control value was based on the original unperturbed value of m, and so we get 0 0 1 . (k+h) m.g+(k+h)x b q (t) = e q(t) + m+m m+m m+m Thus at equilibrium m.g+(k+h)xe (k+h)x = m+m m+m and so the new equilibrium length is m.g x = xe + k+h Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 27 The feedback with gain h has served to reduce the m.g m.g perturbation from to , so that the larger k k+h is h the smaller is the effect of the change of load. Feedback reduces the effect of small disturbances to the point where they may be negligible. Unless h is inordinately large, a large disturbance will yield a significant perturbation. In this case, the brain must, essentially, recompute u(xe) using an estimate of the actual load m+m rather than the "standard load" m. This is our explanation of the long latency reflex noted by Navas and Stark. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 28 What is the price we pay for using a high value of h? We must examine the stability of the new system. We replace k in the system matrix by k+h, which is still positive, and possibly perturb the value of m. Thus the system remains stable. However, for fixed b, k, and m, we can make b2 < 4(k + h)m by making the "feedback gain" h large enough. This happens as soon as h > (b2 /4m) - k. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 29 Theorem: For the mass-spring model with " " coactivation with u(t) = u(xe) and feedback term h(x xe) with gain h for desired equilibrium length xe: (i) the perturbation of the equilibrium length occasioned by a change in load m will be reduced by m.g m.g the feedback from to ; k k+h (ii) the system will be stable; but (iii) even if the system without feedback exhibits no oscillations (b2 > 4km), the feedback system will exhibit oscillations for sufficiently high gain, h > (b2 /4m) - k. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 30 This model can only be part of the story: • Our model is only valid for a range of x in which the linear model of the spring is valid; while the spindle can only monitor x-xe when this value is positive. • A complete model would take into account those muscle spindles which monitor velocity, and the way in which the Golgi tendon organ monitors force. • The efficacy of feedback depends crucially on whether or not the correction signal arrives while it is still valid, and thus delays in feedback paths can have a great effect on the validity and stability of a control system. • Nonlinearities may prove crucial in understanding muscle tremor in terms of limit cycles rather than the harmonic oscillations of a linear system. Michael Arbib CS564 - Brain Theory and Artificial Intelligence, USC, Fall 2001. Lecture 19. Systems Concepts 31