Global parameter identification in systems with a sigmoidal activation function Aleksandar Kojić and Anuradha M. Annaswamy Adaptive Control Laboratory, Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139; email: alko@mit.edu, aanna@mit.edu Corresponding author: Prof. A.M. Annaswamy, Rm 3-461b, MIT, 77 Mass. Ave, Cambridge, MA 02139 Abstract Parameter identification in a 2-node network with sigmoidal activation functions is considered. Given the nonlinearity in the weights, standard estimation algorithms based on linear parametrization are inadequate tools for studying global parameter convergence. In this paper, we provide an alternative approach for studying parameter identification in the presence of sigmoidal parametrization. Conditions under which a simple back propagation algorithm can lead to global convergence are considered. of a neural network when used for nonlinear control. The main idea behind our approach is to directly address and exploit the distinguishing feature of nonlinear regression in neural networks and derive the underlying convergence and stability properties. In our convergence analysis, we make explicit use of the nonlinear sigmoidal activation function and exploit its properties for establishing global convergence of gradientbased methods in neural network training. We focus on a simple 2-node neural network model to derive the underlying properties. 1 Introduction 2 Main result Neural networks have occupied the attention of the engineering community since their introduction less than a decade ago. Both analytically and in practice, it has been proven that neural network architectures have the powerful ability to approximate a wide class of nonlinear functions and thus can be useful tools in many engineering applications [1, 2, 3]. The use of neural networks in identification and control of engineering systems has been intensely debated. Significant progress has been made regarding the statement of the problem, possible ways in which neural networks can be used for identification and control, and demonstrated in seminal numerical and experimental studies [4, 5, 6]. Several stability results have been derived in the literature concerning the use of identification and control (for example,[7, 8]). However, most of them are local in nature and/or include fairly restrictive conditions under which the stability is valid. It should be noted that despite the local stability nature, the actual demonstration in applications and numerical simulations reports just the contrary: Neural networks indeed serve as powerful numerical computational units that are capable of very good approximations of nonlinear maps and provide complex functionalities of estimation, control, and optimization over a large region of operation. In this paper, we take a first, modest step towards closing this glaring gap of explaining the true scope of operation This work was supported by a grant from the the Ford SRL University Research Program. The system of interest is: where is a scalar function of time, parameters with unknown values, and function given by: (1) and are scalar is a sigmoidal ! #"$ !&% (2) We make the following assumptions regarding the system in (1). (A1) (A2) ( '*),+ '-) for . 0 / . Our goal is to design an estimator that estimates as and , respectively, such that if 1 1 2 < 1 and 43 76 98 2 03;: = '*5)@ 'A 1 1 1 1 )CB ?> 11 < the convergence of 1 to 6 is guaranteed. then for all 1 D E l The following estimator is proposed for such a task: 1% 1 where 1 1 GI F H 1 H o (3) . 0 / (4) F 1 1 1 " 1 (5) 5 In what follows, we assume that is a time-varying switching sequence, which is defined as follows: 5 Definition 2.1 A time varying function switching sequence function between and K J , (i) is called a if , there exists a , N M L ' and N L 5N , there exists a , (iii) for any such that PO that PO5 . RQ ' such Q O 5 Q (iv) for each , such that , there that for all 5 T , )VU TX5 WY S , the sets exists a S such 3 Z8 @ : , . \/ are [ W T (>,> > 5 nonempty. (ii) for any such that L that such It can be easily verified that the sigmoidal function in (2) has the following properties )] ) , ^`_ba , 0cKd 5 H 'A)e+ , H fDhg i H W ) , + . H jDhg i (P1) (P2) (P3) An interesting property of symmetry of the estimator can be derived by examining (4). We note that eq. (4) can be rewritten as < symmetric with respect to the set , and therefore we focus all of our discussions that follow on in , where 1 1 % 1 1 GF / C Rk & . 0/ (6) 1 e"$ C k Defining l m= 1 1 1 1 Bn (7) G > l that the trajectories of (6) are symmetric it can be easily seen with respect to in < . Symmetry follows if for any value of the following holds: 1 % 1 1 1 % 1 1 (8) 1 % 1 1 1 % 1 1 (9) From eq. (5) we also have that F 1 1 F 1 1 . Thus, the system is This implies that eqs. (8) and (9) hold. o = 11 1 U 1B (10) > % l The main result that we shall establish in this paper is that all trajectories that begin in <qp converge to 6 . To establish this global result, the following set definitions are required: r r st su s4t v s tv s t v?z r = 1 1 F r 1 1 r ) B (11) r > 8 W W : <qp s t > st(w o s t O st p s4t v = 1 1 1 1 s&t 1 ' B >x yD = 11 11 s t 1 U B >x yD l We shall s show in <qp conr that all trajectories that begin t t s 6. verge to r , and that the trajectories in converge to t s Noting that defines the boundaries of the set , properties of are crucial in establishing the main result. s t and s u This is carried out in Lemma 2.2. Properties of are stated in Lemmas 2.3 and 2.5. Finally, an aspect of differential geometry that is needed in the main result is stated in Lemma 2.6. Due to space considerations, the detailed proofs of lemmas and sublemmas are omitted here. We first begin with some properties of the estimator in . To derive these, the following quantities are useful. o { q 1 | { |1 |1 1% 1% { 1 { q] 1 q] 1 (12) { 1 ~} {{ | q] q] (13) (14) (15) Lemma 2.1 For the estimator in eqs. (3)-(4), the following properties hold : + 1 o ;D { { H * ' ) I } U UA) , (A) , , (B) H (C) )(U U H UA for . 0/ , and for bounded H1 1 , ; and { { H H U*) , 1 P '-) . (D) 1 > 1 0 P H H 1 > that begin in , however, Trajectories stay in . Since these are of measure zero, the result is ”almost global”. r r r We now proceed to derive properties of . In particular, we shall show that and intersect only at two points, and . This is proved by establishing the shape of in the plane for any given . The properties of and are stated below in Lemmas 2.2(a) and (b). r 2 r < 3 r r r < Lemma 2.2 (a) and are represented by two monotonically decreasing convex curves in . r (b) r 6 . We now introduce a metric which we shall use to derive the convergence properties of the estimator. This is defined as follows. For any , o D G H M1 . 0/ H1 (16) r r Thus, the functions and are two non-negative functions, which are zero on and respectively, r and posr itive everywhere else. Therefore, they can be thought of as a measure of distance of a point to the curves , and , respectively. The functions and represent velocities that are associated with each point in < . As the trajectory moves in < , at each point , it will have a velocity specified either by or , depending on the particular the trajectory is at point . value of the input at the time Properties of and are presented in the following lemmas. Sinces their vary depending upon t , s uproperties 2 whether they are in , or in the neighborhood of , they are classified into three distinct Lemmas r 2.3 –2.5. 2.3 Let M1 1 M o . For . 0/ , Lemma D let 1 1 . If U , then s v t , then 1 0 U 1 U 1 , and (b) if (i) (a) if 1 v z s 1 t D , then 1 ' 1 ' 1 , ;D s v t 1 U-) , (ii) + 1 , 1 jD s u o , 1 1 '*) , (iii) + 1 jD | u s Lemma 2.4 If 1 , then the following holds ;D (17) 1 U*)] . \/ 2 2 Lemma 2.5 Let K9 8! be the 2 neighborhood of point 2 2 U : , with bedefined as K (see two ing a Euclidean distance> metric [9]) between 2 2 points and . For all J K K 3 , there exD ] ist T ' T '-) such that U T ¡ ' T ¢ . ¤£ 8 0/x: and . J £ D % (18) st The following lemma is useful in describing the properties of the system trajectories in . ¥ «ª ¦¨§M© co «ª § , § © © ¥ ¥ ¥ (i) the functions ª and ª are twice differentiable, with bounded derivatives ªy¬ , ª#¬ ¬ , and ªy ¬ , ª# ¬ ¬ ; and (ii) ª#¬ § ' ªy ¬ § for all § such that ª § ­ª § , then (A) ¥ is a set consisting of at most one point ª + ' ª ' ª § and + § U , ª 5® § U 5®ª , § and. (B) § ® , § ® We now establish the main result that 6 is a globally attractive equilibrium point for almost all points in < . switching sequence function beTheorem 2.1 If is aassumptions tween and , 5under (A1)-(A2), the solutions 1 of (3)-(4) satisfy the following: l l 5 + ). (i) if 1 ) , then 1 yD <qp l eD 7¯ (ii) if 1 ) , then 1 converges to 6 as ¡Z° . yD ] Proof: l The proof of . is immediate since + 1 , we have ± D 1 1 ²H , and it follows that 1 % 1 % . that H l 1H l H 1 that 1 ). Hence, 1 ) implies for all ] yD CyD #¯ (ii) To prove .. , we proceed in the following three steps. l either converge to 6 as Step 1: Trajectories starting in <qp t s ¡Z° or enter (see figure 1) in a finite time ³ . st is invariant. Step 2: st converge to 6 . r Step 3: Trajectories that start in r For ease of exposition, Figure 1 shows the nature of , , s t , and s u in < for 1 . h D g i Proof of Step 1: We Step 1 bys showing t p 8 f 2 that allKtrajecs u prove tories that start in enter the set 3 : in a finite time. Since is arbitrary, this implies that Step u s 1 holds. Suppose that the trajectories always remain in . Since and are decreasing along the system trajectoU T, ries, there exists a time instant such that 1 ] 5 ' T . If trajectories 1 and hence from Lemma su , 2.5, and hence, there exalways remain in is decreasing, such that 1 U T . But, this implies ists a ' s u , ' T . This is a]5contradiction, that 1 since in 5 as well. is decreasing ¥ and be two curves in the Lemma 2.6 Let ordinate system described by and respectively. Let . If Proof of Step 2: Let st . 1 ] yD .. 1 su at ' 5 eD (19) % From Lemma 2.3 .. and ... , we have that M1 !1 U ) C 1 C1 ' ) % and must have changed sign over the interval ´ µ 0/ . Hence, . D ´ (20) ]1 ) for some D ¶µ s u for ´ µ . Let (20) be true for . . and 1 fD D it follows From](16),(19), andn(20) that there exists a T '­) This implies that for some such that . 1 ) 5 .. 1 ' T (21) 5 s u % which in turn implies that increases in . This con ' tradicts sLemma 2.4. Hence no exists such that u , proving Step 2. 1 ] yD Proof of Step 3: We now show that trajectories starting at st converge any point to 6 .l Since the system beD havior is symmetric withs respect to , we examine only the t v , and show that they converge to trajectories that start in 2 . From Lemma 2.4, it follows r that ifr is kept constant either at or at for all ³ , then the trajectories will converge to a point either in·¯ , or in , respectively. We define the solution of Eq. (6) for , an initial value s t v , and for all ¹ and limit 1 as 1 ¹ 1 ¸D and as ¯º ] 5 points 5 5 » ^b_ba 1 ¹ ¢ 1 5 ¢ . \/ (22) cnd and denote their individual components as ´ 1 0 1 µ½¼ , ´ 1 1 0 µ½¼ . Let the 5 sets ¾ 5 and ¾5 represent the sequence 5 of points in s t v which the5 trajectory converges to and , along respectively, for and . That is, 5 ¾ = s4t v » ^b_ba 1 ¹ BK 5 > D 0cn/ d ] . (23) % 2 It is worth noting that 1 monotonically approaching v s t in is equivalent to 1 increasing and 1 decreas ing montonically with , as well as 1 increasing . We show and 1 decreasing monotonically with 2 s tv that trajectories starting in approach by showing that 1 increases and 1 decreases for . 0/ as ¡¿° . In a s vz manner, it 2 can be shown that all trajectories starting similar t approach . Due to the symmetry of the problem, in st it then would follow that all trajectories starting in O approach 3 . 2 s t v is established Convergence of trajectories to in by considering the two mutually exclusive and collectively s t vz . Us1 t v s t v s t vand exhaustive cases: (b) 1 r (a)r s À z ºD that s t v ing the definitions of D ands t v , s itt vfollows z s t v s t vz ÂÁ , and that 2 and v à v Ä s t and s t . is a limit point for both s t v : In this case, it can be shown that 1 Case (a) 1 s t v . D and 1 0 are decreasing functions of time. Let 1 ]5 eD r r ¾ 5 Å D 5 Æ D + 1 1 ] { ¾ + 5 y¯- 5 K¯~ ¾ ¾ { 1 { { { 5 5 ª 1 ª 1 ¾ ¾ st { { ª#¬ 1 5 ª#¬ 1 5 U3 ª ¬ 1 U ª ¬ 1 + 1 such that ª 1 Ǫ 1 (24) ª ª That is, 2 and satisfy the conditions of Lemma 2.6, and hence and 3 of Lemma 2.6 hold. Suppose that at s t v . From the defi , 1 ÉÈ time instant 5 ® Ê D nition of a switching sequence, there exists a time interval ³ ´ µËYÌ ¾ . Thus, on the interval ³ the trajecv s t . Since 1 tory moves along , from Lemma Í D U 2.3- . - Î , we have that 1 1 . From Lemma 2.1 ¾ increasing with 1 , and that is monotonically we have % thus we have that 1 UY) . Therefore, for all ³ , ED 1 % G 1 H 1 'A) . Hence, on ³ , 1 ' . | H 1 ® have that From Lemma 2.6 3 we ª 1 U ª 1 + ³ (25) 5 #D % , and let 1 Let switch from to at ]5 sequence, Ï Ï 5È Ï . From the property of the switching ´ ¨ µ Ð Ë Ì 5 ® . Pro' such that ³Ï there exists a we note that from Lemma 2.3- . -Î and ceeding as above, s t v . Hence, Lemma 2.1 we have that 1 '*)+ 1 D (26) 1 U Ï + ³Ï ® ¾ 7D % Let ª Ï 1 denote the curve which represents the of 1 for ³Ï . Define 5 a curve ª Ñ 1 as trajectory C yD ªÒÑ 1 ­ª 1 ª Ï ª Ï (27) É® 5® From the definition of Ï and ª Ï and eq.(27), we have ® % that ªÒÑ 1 ­ª Ï ­ª Ï Ï (28) 5® É® % and , and Then, it follows that moves along if , that and along if . From the definition of in (12), it follows that the slope of the tangents to the curves and at any are given by and , respectively. Since and are bounded, it foland which can lows that there exist functions represent the curves and in ,respectively, where the slopes and are given by and , , Lemma 2.1 implies that respectively. Since Also, from (25) and (28), it follows that ªÏ Ï U ª Ï 5® 5® % Since { Ï ª Ï Ï { É® Ï ª É® Ï É® É® it follows from Lemma 2.1 and (29) that 5ÓÔ ª ¬ Ï U ª Ï ¬ Ï 5® 5® % ª Ϭ Ï ª ¬ 5® Ï 5® (29) (30) (31) ª#Ѭ 1 Ǫy¬ 1 1 Differentiating (27) with respect to , we obtain that . Eq. (31) implies that ª Ñ ¬ Ï U ª Ϭ Ï 5® 5 ® % h− D S SE (32) A Eqs. (32) and (28) imply that Lemma 2.6 holds. Therefore, using Lemma 2.6 and (26) it follows that S θ2 ª Ï 1 U ª Ñ 1 + ³Ï 5 7D h+ D (33) SD and hence ª Ï 1 U ª 1 + ³Ï (34) 5 #D % ¾ Since is the limit point of , and is ¾ it follows that that 51 U 1 . a limit point of5 arbitrary , , and 5 , and hence it This relation holds for follows that 1 is a decreasing function of time on ³ . A similar analysis can be carried out to conclude that 1 is a decreasing function of time on ³ Ï . v z s t Case (b): 1 : Similar steps and arguments as preD can be used to conclude that in this case sented in casej(a) 1 and 1 are increasing functions of time.s v It is interesting to note that, because the set t is invariant and bounded, 1 0 and 1 remain bounded in both cases (a) and (b). Since a_bÕ 1 1 + s t v D 1 R a;ÖM× 1 1 + s t vz D 1 it then follows that when is a 2 switching sequence, the system wills converge to the point starting from anywhere in tv . the set s u , the system In summary, for any starting point in t s converges in finite time to . Starting from any point in 2 st , the system converges asymptotically to either point or 3 . Thus, the theorem is established. l Theorem r 2.1 shows that in <qp , all trajectories of (4) converge to 6 . This was established by making use of the which enabled in turn to show that all properties of to s t and that trajectories in s t contrajectories converge verge to 6 . At first glance, the problem solved here appears deceptively simple. Parameter convergence in a nonlinearly parameterized system (1) has been established for very low dimensionality, i.e. . It might then be questioned as to why such elaborate discussions and numerous lemmas and sublemmas were needed to prove this seemingly simple result. We submit that these discussions and properties were essential for proving the main result. The complexity mainly begins with the nonlinear dependence of in . The first implication of is symmetry; both and are stable equilibrium points of the system in (8), and represents a separatrix. As a result, a simple approach that consists of a quadratic positive definite function 9Dmg i 2 l 3 M 1 M SE B 2 θ1 Figure 1: Phase plot for a two-parameter system with sigmoidal Ø parameterization using two different values of . Ù and showing that its derivative is sign-definite is ruled out. The nonlinear dependence of on also implies the existence of manifolds with which are associated regions with certain invariant properties. What has been carried out above is precisely the characterization of these manifolds, given by , , and , and invariant regions given by . s&t l r r References [1] P. J. Werbos. Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences. PhD thesis, Harvard University, 1974. [2] D.E. Rumelhart, G.E. Hinton, and R.J. Williams. Learning representations by back-propagating errors. Nature, 323:533–536, 1986. [3] J. Park and I. W. Sandberg. Universal approximation using radial-basis function networks. Neural Computation, 3:246–257, 1991. [4] K.S. Narendra and K. Parthasarathy. Gradient methods for the optimization of dynamical systems containing neural networks. IEEE Transactions on Neural Networks, 2:252–262, 1991. [5] R.M. Sanner and J-J.E. Slotine. Gaussian networks for direct adaptive control. IEEE Transactions on Neural Networks, 3(6):837–863, November 1992. [6] G. V. Puskorius and L. A. Feldkamp. Neurocontrol of nonlinear dynamical systems with Kalman filter trained recurrent networks. IEEE Transactions on Neural Networks, 5(2):279–297, 1994. [7] S. Yu and A.M. Annaswamy. Stable neural controllers for nonlinear dynamic systems. Automatica, 34:669–679, May 1998. [8] S. Jagannathan, F. L. Lewis, and O. Pastravanu. Discretetime model reference adaptive control of nonlinear dynamical systems using neural networks. International Journal of Control, 64(2):217–230, 1996. [9] Walter Rudin. Principles of Mathematical Analysis. McGraw-Hill, Inc., 1976.