Robot motion classi$cation from the standpoint
of learning control
Shaw-Ji Shiah
a, Kuu-young Young
b;∗aInstitute of Information Science, Academia Sinica, Taipei, Taiwan
bDepartment of Electrical and Control Engineering, National Chiao-Tung University, 1001 Ta Hsueh Road,
30050 Hsinchu, Taiwan
Received 15 February 2001; received in revised form 21 January 2003; accepted 17 February 2003 Abstract
In robot learning control, the learning space for executing the general motions of multi-joint robot ma-nipulators is very complicated. Thus, when the learning controllers are employed as major roles in motion governing, the motion variety requires them to consume excessive amount of memory. Therefore, in spite of their ability to generalize, the learning controllers are usually used as subordinates to conventional controllers or the learning process needs to be repeated each time a new trajectory is encountered. To simplify learning space complexity, we propose, from the standpoint of learning control, that robot motions be classi$ed ac-cording to their similarities. The learning controller can then be designed to govern groups of robot motions with high degrees of similarity without consuming excessive memory resources. Motion classi$cation based on using the PUMA 560 robot manipulator demonstrates the e9ectiveness of the proposed scheme.
c
2003 Elsevier B.V. All rights reserved.
Keywords: Robot motion classi$cation; Robot learning control; Learning space complexity; Motion similarity analysis
1. Introduction
The dynamics of robot manipulators are, in general, non-linear and complex. Therefore, conven-tional $xed gain, linear feedback controllers are not capable of e9ectively controlling the movements of multi-joint robot manipulators under di9erent distance, velocity, and load requirements. Through the use of non-linear feedback, approaches like the computed torque method provide better com-pensation for the dynamic interactions present in various robot motions [15]. But, these approaches demand complete, non-linear dynamic models describing the robot manipulator, which are diAcult to
∗Corresponding author. Tel.: +886-3-5712-121; fax: +886-3-5715-998.
E-mail address: kyoung@cc.nctu.edu.tw(K.-y. Young).
0165-0114/$ - see front matter c 2003 Elsevier B.V. All rights reserved.
be accurately modeled and implemented in real-time. On the other hand, learning controllers, such as neural networks and fuzzy systems, are attractive alternatives in robot motion control, because they are able to tackle highly complex dynamics without explicit model dependence and identi$cation, in addition to their capability in generalization [5,9,13,19]. However, learning controllers are usually used as subordinates to conventional controllers in governing robot motions [8,11]. The conventional controller is responsible for the major portion of the control, and brings the system close to the desired state, after which the learning controller compensates for the remaining error. Some learning control schemes do use learning controllers alone to execute motion control [6]. But, most of these schemes need to repeat the learning process each time a new trajectory is encountered. Otherwise, a neural network will consist of a huge number of neurons or a fuzzy system will require too many rules. This learning controller de$ciency results mainly from the complexity of motions associated with various task requirements. Consequently, when a learning controller is given a major role in governing the general motion of a multi-joint robot manipulator, the learning space it must deal with is extremely complicated [14,17,20].
To simplify the complexity of the learning space in using learning controllers to govern robot motions, we propose, from the standpoint of learning control, that robot motions be classi$ed ac-cording to their similarities. Thus, learning controllers can then be designed to govern groups of robot motions with high degrees of similarity with smaller memory sizes. By contrast, when robot motions are randomly arranged, learning controllers will demand larger memory sizes in motion governing. For instance, in the authors’ previous paper [21], we developed a robot learning control scheme that generalizes the parameters of the fuzzy systems, which are appropriate for the govern-ing of the sampled motions in a class of motions, to deal with the whole class of motions. Then, when the motions in the class are with high degrees of similarity, the learning control scheme can govern the class of motions with a small memory size. Thus, more robot motions can be governed by the scheme, with a $xed memory size, when they are grouped into classes of similar motions appropriately. In this study, we use a fuzzy system to perform motion similarity analysis and classi-$cation. When the fuzzy system learns to govern motion successfully, similarities between motions are evaluated by analyzing the fuzzy parameters in the fuzzy system. The rest of this paper is orga-nized as follows. The proposed motion similarity analysis and classi$cation and its implementation are discussed in Section 2. In Section 3, simulations based on the use of a two-joint planar robot manipulator and the PUMA 560 robot manipulator are reported. Discussions and conclusions are in Section 4.
2. Motion similarity analysis and classication
Motion similarity can be de$ned according to di9erent characteristics [12,16]. For example, a number of arbitrary robot motions can be categorized into classes of motions with similar movement distances, velocities, or loads [21]. However, this classi$cation cannot guarantee that motions in the same class will correspond to similar fuzzy parameters when governed using a fuzzy system. In the proposed approach, we aim to group similar motions to simplify the complexity in the learning space. Therefore, from the standpoint of learning control, we take similarities between motions as similarities between the fuzzy parameters of the governing fuzzy systems, and de$ne motion similarity as follows.
Motion Governing Using Simplified FNNs Motion Classification Motion Similarity Measurement Motion Governing Using FNNs Arbitrary motions FNN parameters Degree of motion similarities Classified similar motions
Fig. 1. Conceptual organization of the proposed robot motion similarity analysis and classi$cation.
Denition 1 (Motion similarity). Two motions governed using the fuzzy system are said to be simi-lar if the fuzzy parameters of the governing fuzzy systems, i.e., the fuzzy rules and input and output membership functions, are similar.
According to De$nition 1, Fig. 1 shows the conceptual organization of the proposed motion similarity analysis and classi$cation. In Fig.1, arbitrary input motions are $rst governed using a fuzzy neural network (FNN). The FNN, discussed in Section 2.2, is basically a fuzzy system implemented using a neural network structure, so that the fuzzy parameters can be adjusted automatically [1,4,9]. Initially, a large number of FNN linguistic labels are used in the learning. The learning process will terminate when the FNN can successfully govern the motions up to a pre-speci$ed accuracy. During learning, redundant fuzzy rules in the FNN are eliminated. The resultant fuzzy parameters are then evaluated via the process of motion similarity measurement. Thus, according to the degrees of similarity between these fuzzy parameters, the motions input in arbitrary fashion are classi$ed into groups of similar motions, which can then be governed using simpli$ed FNNs.
In evaluating the similarities between the fuzzy parameters of the governing FNNs, it is quite straightforward to compare the numbers of fuzzy rules and the shapes of the corresponding mem-bership functions in the FNNs. In the authors’ previous paper [22], we de$ned FNN similarity as follows.
Denition 2 (FNN similarity (I)). Two FNNs for motion governing are said to be similar if the numbers of fuzzy rules they possess are the same, and the similarity among the shapes of their corresponding membership functions is above a pre-speci$ed threshold.
De$nition 2 is very strict, because of the restriction on the number of fuzzy rules in the FNN. In addition, it takes individual checking in evaluating the similarities between the membership functions corresponding to the fuzzy rules. Thus, by comparing two fuzzy systems as a whole through eval-uating the fuzzy relations representing the entire fuzzy systems, in this paper, we propose another de$nition of FNN similarity as follows.
Denition 3 (FNN similarity (II)). Two FNNs for motion governing are said to be similar if the fuzzy relations representing the characteristics of the FNNs are similar.
In De$nition 3, the number of fuzzy rules is disregarded in the FNN similarity evaluation. Because De$nition 3 is less restrictive than De$nition 2, two FNNs tend to be determined as similar under De$nition 3. Later, in Section 3, we evaluate the e9ects of these two de$nitions via simulations. 2.1. System implementation
In this section, we discuss how to implement the proposed motion similarity analysis and clas-si$cation according to De$nition 3. System implementation according to De$nition 2 can be found
in [22]. Assume that FN1 and FN2 are two FNNs with two inputs and one output, and they govern
Motion 1 and Motion 2 well with N1 and N2 fuzzy rules, respectively. Take FN1 as an example. Let
x ∈ NX1 and y ∈ NY1 be two non-fuzzy input variables representing the position and velocity of some
joint of the robot manipulator, and z ∈ NZ1a non-fuzzy output variable representing the command sent
to the robot manipulator, where NX1; NY1; NZ1⊂ R. Letting F(·) represent a fuzzy set, and X1∈ F( NX1),
Y1∈ F( NY1), and Z1∈ F( NZ1) be the linguistic variables representing the two fuzzy input variables and
one fuzzy output variable, respectively, the fuzzy rules in FN1 can then be expressed as
If X1 is A11 And Y1 is B11 Then Z1 is C11;
If X1 is A12 And Y1 is B12 Then Z1 is C12;
· · ·
If X1 is A1N1 And Y1 is B1N1 Then Z1 is C1N1; (1)
where A1i, B1i, and C1i, i = 1; : : : ; N1, are linguistic values of X1, Y1, and Z1, respectively. Let
R1∈ F( NX1× NY1× NZ1) be the fuzzy relation representing FN1. We can then express R1 as
R1= {((x; y; z); R1(x; y; z)) | (x; y; z) ∈ NX1× NY1× NZ1} (2) with
R1(x; y; z) = sup
i min(A1i(x); B1i(y); C1i(z)); (3)
where F(·) : U → [0; 1] stands for a membership function characterizing a fuzzy set F [7,10].
Similarly, the fuzzy relation R2 representing FN2 can be expressed as
R2= {((x; y; z); R2(x; y; z)) | (x; y; z) ∈ NX2× NY2× NZ2} (4) with
R2(x; y; z) = sup
i min(A2i(x); B2i(y); C2i(z)); (5)
where A2i, B2i, and C2i are linguistic values of X2, Y2, and Z2, respectively. With R1 and R2, we
de$ne the similarity index, ∈ (0; 1), between FN1 and FN2 for governing Motions 1 and 2 as
SM(R1; R2) = ; (6)
where SM(· ; ·) is a similarity measurement operator. Because R1 and R2 are in the form of fuzzy
sets, the similarity evaluation between R1and R2using the operator SM can be realized by evaluating
the similarity between the fuzzy sets corresponding to R1 and R2.
The similarity measurement between two fuzzy sets U1 and U2, SM(U1; U2), can be de$ned as
SM(U1; U2) = M(UM(U1∩ U2)
1∪ U2); (7)
where ∩ and ∪ denote the intersection and union operators, respectively, and M(·) is the size of a fuzzy set. The two famous methods to measure the similarity between fuzzy sets are the geometric and set-theoretic measures [24]. For the geometric measure, similarities between fuzzy sets are
computed by comparing the areas covered by the fuzzy sets according to geometric points [2,9]. In using the geometric measure, the fuzzy sets need to be normalized to place the similarity evaluation on the same scale, because various input motions may correspond to di9erent ranges of movement distances and velocities [22]. To avoid the normalization process in similarity evaluation, we adopted the set-theoretic measure and describe the procedure as follows.
Take FN1 as an example. We $rst sample the spaces, NX , NY, and NZ, with n equally-spaced points,
and discretize the fuzzy sets, A1i, B1i, and C1i into ˆA1i, ˆB1i, and ˆC1i, described as
ˆ
A1i= {(xr; A1i(xr)) | r = 1; 2; : : : ; n}; (8)
ˆB1i= {(ys; B1i(ys)) | s = 1; 2; : : : ; n}; (9)
ˆC1i= {(zt; C1i(zt)) | t = 1; 2; : : : ; n}: (10)
By using Eqs. (8)–(10), the membership function R1(xr; ys; zt) can be derived as R1(xr; ys; zt) = sup
∀i min(A1i(xr); B1i(ys); C1i(zt)): (11)
Let ˜xj= (xr; ys; zt)j, with j = 1; 2; : : : ; n3. The discretized fuzzy relation R1 can then be expressed as
ˆR1= {( ˜xj; R1( ˜xj)) | j = 1; 2; : : : ; n3}: (12)
Similarly, the discretized fuzzy relation R2 can be expressed as
ˆR2= {( ˜yj; R2( ˜yj)) | j = 1; 2; : : : ; n3}: (13)
Finally, the similarity between ˆR1 and ˆR2 is evaluated using SM( ˆR1; ˆR2), described in Eq. (14) [3]:
SM( ˆR1; ˆR2) =| ˆR| ˆR1∩ ˆR2| 1∪ ˆR2| = n3 k=1min(R1( ˜xk); R2( ˜yk)) n3 k=1max(R1( ˜xk); R2( ˜yk)) ; (14)
where | · | is the cardinality operator [23]. 2.2. The FNN learning mechanism
The FNN learning mechanism used in this paper is shown in Fig. 2. The representation of a fuzzy system using a fuzzy neural network enables us to take advantage of the learning capabil-ity of the neural network for automatic tuning of the parameters in the fuzzy system. The fuzzy reasoning parameters are thus expressed in terms of the connection weights or node functions of the neural network [1,4,9,18]. We chose an FNN with a structure similar to that in [9], of course, other types of FNN can also be used. As Fig. 2 shows, the inputs to the FNN are position and velocity trajectories of input motions, qri and ˙qri, and the outputs are motion commands Cmi. There
... mi C ri q ri q Min operation Defuzzification (COA) Fuzzification Layer1 (Input nodes) Layer2 (Input membership nodes) Layer3 (Rule nodes) Layer5 (Output node) Layer4 (Output membership nodes) Rule strength transmission .
Fig. 2. The structure of the FNN.
the output membership layer, and the output layer. Gaussian functions with adjustable means and variances were used as membership functions. A gradient-descent-based back-propagation algorithm was employed for learning [6]. During the learning process, a large number of FNN linguistic labels were initially chosen in arbitrary fashion and normal fuzzy sets were used as membership functions. The learning process terminated when the FNN could govern motion successfully; i.e., the position mean square error was less than a pre-speci$ed value. After the input motion had been learned, the similarities between membership functions corresponding to this motion were evaluated pair by pair. When membership functions were very similar, it indicates that some of the linguistic labels were unnecessary and could be eliminated. Therefore, after the learning process, the FNN would have a simpli$ed structure and be ready for similarity measurement between motions.
3. Simulation
Simulations were performed to demonstrate the e9ectiveness of the proposed motion similarity analysis and classi$cation based on the use of a two-joint planar robot manipulator and the PUMA 560 robot manipulator. The dynamics of multi-joint motions can be formulated as follows:
Table 1
The kinematic and dynamic parameters for the PUMA 560 robot manipulator
Link Link mass (kg) Inertial matrix (kg m2) Center of mass (m)
Ixx Iyy Izz x y z Dynamic parameters 1 17.085 0.661133 0.661133 0.098877 0 0 −0:08 2 39.423 0.365234 3.577148 3.711426 −0:216 0 −0:0675 3 18.513 0.381836 0.393555 0.06665 0 0 0.216 4 4.5645 0.012695 0.009521 0.012695 0 −0:02 0 5 1.2189 0.0007324 0.0014648 0.0007324 0 0 0 6 0.51 0.001709 0.001709 0.001 0 0 0 Joint i i (deg) ai di Kinematic parameters 1 1 90 0 0:671 m 2 2 0 0:432 m 0:15 m 3 3 −90 0:02 m 0 4 4 90 0 0:433 m 5 5 −90 0 0 6 6 0 0 0
where q, ˙q, and Tq stand for joint variables and their derivatives, H(q) is the inertia matrix, C(q; ˙q) is the vector of centrifugal and Coriolis terms, G(q) is the vector of gravity terms, and B is the vector of joint torques. The e9ect of gravity was ignored in the simulations. The kinematic and dynamic parameters for the two-joint planar robot manipulator are: link length, l1= 0:30 m
and l2= 0:32 m, link mass, m1= 2:815 kg and m2= 1:640 kg, center of mass, lc1= 0:15 m and
lc2= 0:16 m, and inertia, I1= I2= 0:0234 kg m2; those for the PUMA 560 robot manipulator are
listed in Table 1. To provide various input motions, a second-order system was used, as described below:
L T + B ˙ + K( − d) = 0; (16)
where L is the load, K the sti9ness, B the damping coeAcient, and and d the actual and desired
joint positions for each joint, respectively. Di9erent motions were generated by varying L, B, K, and d. Each joint of the robot manipulator was equipped with an FNN. The inputs to the FNN were
the position and velocity trajectories of the input motions, and the output was the motion command. Fifty equally-spaced points were used for the discretization of the fuzzy sets represented in the FNN. Similarities between the motions were evaluated according to the values of the similarity indices as
= min
16l6ln l = min
Fig. 3. A group of motions executed using a two-joint planar robot manipulator.
where ˆRpl and ˆRql were the discretized fuzzy relations of the FNNs which governed the lth joint of
the robot manipulator for Motions p and q, and ln equal to 2 and 6 for the two-joint planar robot
manipulator and the PUMA 560 robot manipulator, respectively.
In the $rst set of simulations, we applied the proposed approach according to De$nitions 2 and3, respectively, to analyze the similarities between the group of motions shown in Fig. 3, which were executed using the two-joint planar robot manipulator. The motions in Fig.3 were generated to start from the same position and reach di9erent end positions with L, B, and K in Eq. (16) being the same. Because these motions were generated under very similar kinematic and dynamic conditions, they were expected to be determined as similar by using the proposed approach. Table 2 shows the degrees of similarity between motions in Fig. 3 according to De$nitions 2 and 3, respectively. In Table 2, high degrees of similarity between these $ve motions were observed under both de$nitions, and De$nition 2 led to higher degrees of similarity. In Table 2, we also observed that the degrees of similarity under the analysis according to De$nition 3 monotonically decreased along with the increase of the distances between these $ve motions. This phenomenon was not present in the similarity analysis according to De$nition 2. The results implicate that FNN similarity evaluation according to De$nition 3 seems to correspond to the closeness of the motions in distance, while further investigation is demanded for solid conclusions.
In the second set of simulations, we intended to evaluate how the increase of joints in the robot manipulator would a9ect the e9ects of De$nitions 2 and3. According to both de$nitions, we applied the proposed approach to analyze the similarities between the group of motions shown in Fig. 4, which were executed using the PUMA 560 robot manipulator. The simulation results show that all motions are determined as dissimilar when De$nition 2 was used. It is because high variations
Table 2
The degrees of similarity between motions in Fig. 3according to (a) De$nition2and (b) De$nition3
1 2 3 4 5 (a) De6nition 2 1 1 0.963 0.966 0.943 0.921 2 1 0.927 0.938 0.962 3 1 0.975 0.953 4 1 0.917 5 1 (b) De6nition 3 1 0.834 0.718 0.587 0.484 1 0.842 0.680 0.554 1 0.784 0.622 1 0.763 1
Fig. 4. A group of motions executed using the PUMA 560 robot manipulator.
were present in the rule numbers and the corresponding membership function distributions of the FNNs governing the motions executed by the six-joint PUMA 560 robot manipulator. On the other hand, some motions were still classi$ed as similar under the less strict De$nition 3. Table 3 shows the degrees of similarity between motions in Fig. 4 according to De$nition 3, and Table 4 the classi$cation of motions in Fig. 4 according to di9erent values of similarity indices . The results demonstrate that De$nition 3 yielded better performance than De$nition 2 when the six-joint robot manipulator case was involved.
In Table 4, we found that Motions 1 and 2 and Motions 5 and 6 have similarity index values higher than 0.7. In the third set of simulations, we intended to show that fuzzy parameters for governing motions with high degrees of similarity could be generalized to govern similar motions.
Table 3
The degrees of similarity between motions in Fig. 4according to De$nition3
1 2 3 4 5 6 1 1 0.739 0.422 0.446 0.073 0.073 2 1 0.455 0.475 0.072 0.073 3 1 0.395 0.069 0.069 4 1 0.071 0.071 5 1 0.808 6 1 Table 4
Classi$cation of motions in Fig.4according to their similarities
N Motion classes Number of classes
0.9 (Motion 1), (Motion 2), (Motion 3), (Motion 4), (Motion 5), (Motion 6) 6 0.7 (Motions 1, 2), (Motions 5, 6), (Motion 3), (Motion 4) 4 0.5 (Motions 1, 2), (Motions 5, 6), (Motion 3), (Motion 4) 4
0.3 (Motions 1, 2, 3, 4), (Motions 5, 6) 2
Fig. 5. Motion governing by using the FNN with generalized fuzzy parameters.
We $rst performed simulations for Motions 1 and 2, which were two similar motions with loads equal to 0 and 5 kg, respectively. We generalized the fuzzy parameters for the FNNs governing these two motions to govern similar motions with loads ranging between 0 and 5 kg. Fig. 5 shows the result when the load was equal to 2:5 kg, and the generated motion approximates the reference motion quite well. Similar results were observed for other loads. We also performed simulations for Motions 5 and 6, and the results were similar to those for Motions 1 and 2. Thus, we concluded that, via the proposed motion similarity analysis, motions may be classi$ed as similar, and these
similar motions can then be governed using generalized similar fuzzy parameters, implicating that the learning controller can be designed to govern more motions with smaller memory allocation. 4. Discussion and conclusion
This paper has proposed motion similarity analysis from the standpoint of learning control. Similar motions were de$ned as those corresponding to similar fuzzy parameters when governed using fuzzy systems. By classifying motions according to their similarities, learning controllers can be designed to govern groups of motions with high degrees of similarity with smaller memory sizes. Simulations based on the use of the PUMA 560 robot manipulator veri$ed the e9ectiveness of the proposed scheme.
From the simulation results in Section 3, we can $nd that the motions might be categorized into di9erent motion groups, when di9erent similarity indices were chosen for motion similarity evaluation. With a larger (smaller) similarity index, the motions in the same group may be more similar (dissimilar); consequently, the fuzzy parameters of the FNNs for governing these motions can be generalized to govern other similar motions with higher (lower) precision. Thus, similarity index selection may depend on the demanded accuracy in motion governing using the generalized fuzzy parameters.
A point that also deserves discussion is about the e9ects of adopting di9erent types of FNNs for the proposed scheme. It can be expected that when di9erent types of FNNs were used for similarity analysis, the resulting analysis and subsequent motion classi$cation might be somewhat di9erent. However, we consider which types of FNNs to be used in the proposed scheme may not be that crucial, if only the motions can be classi$ed into groups of motions with high degrees of similarity and governed by using learning controllers with smaller memory allocation.
In future works, we will apply the proposed scheme to classify general robot motions over the entire learning space, so that an organized and simpli$ed learning space for motion governing may be achieved. Simulation results in Section 3 demonstrate that motion classi$cation via the means of learning does not necessarily correspond to the kinematic or dynamic features, and further investiga-tion into similarities among the general moinvestiga-tions is then demanded. In addiinvestiga-tion, the proposed scheme will also be utilized for practical applications.
Acknowledgements
This work was supported in part by the National Science Council, Taiwan, under grant NSC 90-2213-E-009-093.
References
[1] H.R. Berenji, P. Khedkar, Learning and tuning fuzzy logic controllers through reinforcements, IEEE Trans. Neural Networks 3 (5) (1992) 724–740.
[2] C.T. Chao, Y.J. Chen, C.C. Teng, Simpli$cation of fuzzy-neural systems using similarity analysis, IEEE Trans. Systems Man Cybernet. Part B: Cybernet. 26 (2) (1996) 344–354.
[3] D. Dubois, H. Prade, Fuzzy Sets and Systems: Theory and Applications, Academic, New York, 1980.
[4] J.S.R. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans. Systems Man Cybernet. 23 (3) (1993) 665–685.
[5] A. Karakasoglu, S.I. Sudharsanan, Identi$cation and decentralized adaptive control using dynamical neural networks with application to robotic manipulators, IEEE Trans. Neural Networks 4 (6) (1993) 919–930.
[6] M. Kawato, Y. Uno, M. Isobe, R. Suzuki, Hierarchical neural network model for voluntary movement with application to robotics, IEEE Control Systems Mag. 8 (2) (1988) 8–16.
[7] C.C. Lee, Fuzzy logic in control systems: fuzzy logic controller—Part I and II, IEEE Trans. Systems Man Cybernet. 20 (2) (1990) 404–435.
[8] C.M. Lim, T. Hiyama, Application of fuzzy logic control to a manipulator, IEEE Trans. Robotics Automat. 7 (5) (1991) 688–691.
[9] C.-T. Lin, A neural fuzzy control system with structure and parameter learning, Fuzzy Sets and Systems 70 (1995) 183–212.
[10] J.H. Lu, J.Q. Chen, L.J. Chen, Analysis and synthesis of fuzzy closed-loop control systems, IEEE Trans. Systems Man Cybernet. 25 (5) (1995) 881–888.
[11] W.T. Miller III, F.H. Glanz, L.G. Kraft III, Application of a general learning algorithm to the control of robotic manipulators, Internal. J. Robotics Res. 6 (2) (1987) 84–98.
[12] M.C. Nechyba, Y. Xu, Stochastic similarity for validating human control strategy models, IEEE Trans. Robotics Automat. 14 (3) (1998) 437–451.
[13] T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, Y. Uchikawa, Trajectory control of robotic manipulators using neural networks, IEEE Trans. Ind. Electron. 38 (3) (1991) 195–202.
[14] T.D. Sanger, Neural network learning control of robot manipulators using gradually increasing task diAculty, IEEE Trans. Robotics Automat. 10 (3) (1994) 323–333.
[15] L. Sciavicco, B. Siciliano, Modeling and Control of Robot Manipulators, McGraw-Hill, Singapore, 1996.
[16] M. Setnes, R. Babuska, U. Kaymak, H.R. van Nauta Lemke, Similarity measures in fuzzy rule base simpli$cation, IEEE Trans. Systems Man Cybernet. Part B: Cybernet. 28 (3) (1998) 376–386.
[17] T. Shibata, T. Fukuda, Hierarchical intelligent control for robotic motion, IEEE Trans. Neural Networks 5 (5) (1994) 823–832.
[18] C.T. Sun, Rule-base structure identi$cation in an adaptive-neural-based fuzzy inference system, IEEE Trans. Fuzzy Systems 2 (1) (1994) 64–73.
[19] B.A.M. Wakileh, K.F. Gill, Use of fuzzy logic in robotics, Comput. Ind. 10 (1988) 35–46.
[20] K.Y. Young, C.C. Fan, An approach to simplify the learning space for robot learning control, Fuzzy Sets and Systems 95 (1) (1998) 23–38.
[21] K.Y. Young, S.J. Shiah, An approach to enlarge learning space coverage for robot learning control, IEEE Trans. Fuzzy Systems 5 (4) (1997) 511–522.
[22] K.Y. Young, J.K. Wang, Robot motion similarity analysis using an FNN learning mechanism, Fuzzy Sets and Systems 124 (2) (2001) 155–170.
[23] H.J. Zimmermann, Fuzzy Set Theory and its Applications, 3rd Edition, Kluwer, Norwell, MA, 1996.
[24] R. Zwick, E. Carlstein, D.V. Budescu, Measures of similarity among fuzzy concepts: a comparative analysis, Internal. J. Approximate Reasoning 1 (1987) 221–242.
