You are on page 1of 9

Learning Space Analysis and Coverage for Robot Learning Control

Shaw-Ji Shiah and Kuu-young Young


Department of Electrical and Control Engineering
National Chiao-Tung University, Hsinchu, Taiwan
Abstract human beings resolve this learning space problem.
To tackle the nonlinearity present in the dynamics of In our daily life, we may see frequently the cases that
robot manipulators, the learning controllers have been the human motor control system deals with large learn-
applied for robot motion governing using dierent kinds ing space. For example, the arm postures and the forces
of control structures. However, most of them need to required for a basketball player to shoot a ball are dier-
repeat the learning process each time a new trajectory ent along with the locations she (he) is and the actions
is encountered. Otherwise, a neural network will consist that she (he) takes. It is impossible for this player to
of a huge number of neurons or a fuzzy system will re- practice for handling every possible situation in advance.
quire too many rules because the learning space needed Thus, the movements of shooting a ball should be similar
to handle arbitrary trajectories is too large. Inspired by in certain degree so that a basketball player can make
the concept of human motor program, we consider that use of the existing skills to deal with unpracticed situa-
to generalize the learned motions eectively for govern- tions. The concept of human motor program also states
ing those unpracticed motions is the key role to reduce that a motor program should be generalized, simple to
the size of the learning space. In this dissertation, a operate, and eceint in storage 16]. Because of these
novel robot learning control scheme, which consists of a reasons, we consider that to generalize the learned mo-
learning control structure emulating human motor pro- tions eectively for governing those unpracticed motions
grams and includes a robot motion similarity measure- is the key role for reducing the size of the learning space.
ment algorithm, is proposed to analyze learning space In this dissertation, we propose a novel learning scheme
and enlarge learning space coverage. Simulation results that is capable of large learning space coverage. A new
show that the learning eort is dramatically reduced in robot learning control structure is proposed for this learn-
dealing with a wide range of robot motions by using the ing control scheme 23]. It is motivated by the con-
proposed scheme. cept of human motor program, and consists mainly of
a fuzzy system and a cerebellar model articulation con-
troller (CMAC)-type neural network 1, 2]. The fuzzy
1.Introduction system is used for governing a number of sampled mo-
The dynamics of robot manipulators are, in gen- tions in a group of motions which are not yet classied.
eral, nonlinear and complex. Therefore, conventional Meanwhile, we also propose to evaluate the degrees of
xed gain, linear feedback controllers are not capable of similarities between robot motions so as to classify those
eectively controlling movements of multijoint robot ma- motions governed by the fuzzy system 17, 18]. The
nipulators under dierent task requirements. To achieve CMAC is then used to generalize the parameters of the
better compensation for dynamic interaction, a com- fuzzy system to deal with the whole group of motions.
plete, nonlinear dynamic model describing the robot ma- One reason for adopting the CMAC rather than another
nipulator is needed 7]. The use of a complicated nonlin- type of neural network is that this neural network has a
ear dynamic model makes real-time implementation dif- good generalization capability and simple structure. Un-
cult. On the other hand, if a learning controller is used, der this design, in some sense the qualitative fuzzy rules
the consequences of using incomplete models and inaccu- in the fuzzy system are generalized by the CMAC-type
rate model parameters may not be very signicant. Two neural network, and then a larger learning space can be
well-known types of learning controllers for robot motion covered 23]. Furthermore, motion classication accord-
control are articial neural networks and fuzy systems. ing to similarities reduces the complexity of learning.
Both are biologically inspired and intended to model hu- Therefore, the learning eort is dramatically reduced in
man experience 20, 25]. However, most of them need to dealing with a wide range of robot motions.
repeat the learning process each time a new trajectory is
encountered. Otherwise, a neural network will consist of 2.Human Motor Control
a huge number of neurons or a fuzzy system will require
too many rules because the learning space needed to The proposed learning control scheme is motivated
handle arbitrary trajectories is too large 15, 19]. Since by the concept of human motor program, which is de-
the human motor control system can perform numerous rived from the study of human movement and has stim-
kinds of motions through learning, we wonder how the ulated research in human control strategies 3, 5, 9, 10,
14, 16]. Because feedback processing in the human motor fuzzy rules is more eective than that of quantitative
control system is slow, long delays are experienced in the numerical data, because the former involves the general-
transfer of sensory information to the higher level of the ization of abstract representations and tends to cover a
hierarchy. For slow movement, long delays may cause no larger learning space.
serious problems when feedback control is employed by
the higher level of the hierarchy however, the eect of
delays cannot be ignored in dealing with fast movement. 3.1.Robot Learning Control Structure
Therefore, open-loop control might be more appropriate The conceptual organization of this learning sturcture
for governing fast human movement and a concept of is shown schematically in Fig. 1. In Fig. 1, the inputs
motor program was proposed 5, 16]. The basic idea in to the learning structure are motions which belong to
human motor program is that movement demands are a motion class. Each of these motions can be repre-
specied by the CNS in advance and then executed in sented as mi (pi ), where mi refers to the ith motion in
an essentially open-loop manner. When activated, the the motion class, and pi is the corresponding vector of
motor program generates motor commands and sents parameters describing the ith robot motion. For a mo-
them to the peripheral neuromotor system for execution. tion mi , the motion parameter vector pi usually consists
Where a class of movements are concerned, the storage of movement initial position, nal position, distance, ve-
of complex motor programs for each movement may not locity, load, and other kinematic or dynamic constraints.
be appropriate. Instead, the motor program should be The motion parameter vector pi is used as an index and
generalized, simple to operate, and ecient in storage. inputted into the CMAC. A set of parameters Fi will
Thus, a single motor program will be enough to accom- then be solicited from the CMAC and sent forward to the
modate a wide variety of movements. Proper parame- fuzzy system. The kinematic-related elements of pi are
ters corresponding to a particular movement can then sent to the motion trajectory planner, and the planner
be supplied to the motor program in execution. In other generates the corresponding desired joint position and
words, the motor program should be abstract and the velocity trajectories which are used as the inputs to the
corresponding control parameters should be eciently fuzzy system. With Fi representing the fuzzy rules, the
stored and manipulated 13, 16]. fuzzy system can govern the motion along with the de-
sired position and velocity trajectories of the motion. To
3.Proposed Learning Control Scheme allow for automatic adjustment of the system's param-
eters, the fuzzy system in the learning structure is im-
The proposed learning control scheme uses a new plemented in the form of a fuzzy neural network (FNN)
learning control structure 23] and includes a motion sim- 4, 12], and generates motion commands sent forward to
ilarity measurement algorithm 18]. The learning con- a local controller. In turn, the local controller, which
trol structure consists mainly of a fuzzy system and a emulates the peripheral neuromotor system, modulates
CMAC. We take advantage of the merits of a fuzzy sys- the motion command via sensory feedback and uses the
tem and a CMAC to emulate a motor program with the resultant signal to move the robot manipulator 21, 22].
appealing characteristics described in Sec. 2. A fuzzy According to some biological evidence, a simple position
system is used to represent the abstract motor program control law with linear damping is then adopted for the
and a CMAC to manage the parameters. In some sense, local controller 5, 21]
the fuzzy system encodes knowledge via qualitative rules
rather than via precise quantitative description. This  =Kp(Cm ; q) ; Kd q_ (1)
feature makes the fuzzy system an appealing choice for
emulating a motor program. The parameters specifying where Cm stands for the motion command vector, q and
the fuzzy rules for governing sampled motions are stored q_ are the actual position and velocity vectors obtained
and manipulated by the CMAC to deal with a wide range from sensory feedback, and Kp and Kd are selected to
of motions. The combination of a fuzzy system and a be diagonal positive denite gain matrices for stability
CMAC in this way is novel and exploits the merits of considerations 7].
the motor program. In order to simplify the complexity
in learning, we also propose to evaluate the degrees of 3.2.Motion Similarity Analysis
similarities between robot motions so as to classify those
motions governed by the fuzzy system 17, 18]. Under It is quite intuitive that the characteristics of a
this arrangement, a class of motions with the same fea- robot motion for a special control purpose will be re-
ture are expected to correspond to similar fuzzy param- ected on the dynamics of that robot manipulator. There-
eters. Consequently, the data with which the CMAC, fore, to measure the degree of the similarity between
which executes the generalization, will have to deal will robot motions is equal to measure that of the closeness
exhibit less nonlinearity. Generalization of qualitative between the corresponding robot dynamics. As direct
 
  
 
5L
/ L

  2L  6G 6G 2


9 5RERW

 
3L

   


 '\QDPLFV
    
 


6 6

Figure 1: Conceptual organization of the proposed robot learning control structure.

measurement of similarity between robot dynamics is im- where F () represents a fuzzy set X1 2 F (X 1 ), Y1 2
possible, we will measure the similarity between rules or F (Y 1 ) and Z1 2 F (Z 1 ) are the linguistic variables rep-
parameters in the fuzzy system that control the robot resenting two fuzzy input variables and one fuzzy output
motions. Thus, we dene similar motions as follows. variable. A1i , B1i , and C1i dened on F (X 1 ), F (Y 1 )
Denition 1 (Similar Robot Motions) If the fuzzy and F (Z 1 ), are linguistic values of X1 , Y1 and Z1 , where
relations of the fuzzy systems govern a group of robot i = 1 2 : : :  N . Let R1 2 F (X 1  Y 1  Z 1 ) be the fuzzy
motions are similar, these robot motions are similar. relation on FN1 . We can then determine R1 by using
In this dissertation, we choose the fuzzy relations of the each R1i as 6, 11, 24]:
fuzzy system governing robot motions for evaluation. R1 = _Ni=1 R1i = _Ni=1 (A1i ^ B1i ^ C1i ) (3)
Because the fuzzy relation of a fuzzy system describes
the fuzzy system as a whole, it is appropriate to mea- where _ and ^ represent max and min connectives, re-
sure the similarity between fuzzy systems by evaluating spectively 6]. The fuzzy set form can be expressed as
their fuzzy relations. In evaluating the similarities be- R1 =f(x y z ) R1 (x y z ) j (x y z ) 2 X 1  Y 1  Z 1 g
tween fuzzy relations, we dene similar fuzzy relations (4)
as follows. with
Denition 2 (Similar Fuzzy Relations) If fuzzy re-
lations are with similar characteristics or distributions, R1 (x y z )= sup min(A1 (x) B1 (y) C1 (z )) (5)
these fuzzy relations are similar. i i i i

Consider two fuzzy systems implemented as two FNNs, where F (:) : U ! 0 1] stands for a membership func-
FN1 and FN2, with two inputs and one output. Assume tion characterizing a fuzzy set F. This fuzzy set R1 can
that FN1 and FN2 govern motion 1 and motion 2 well, also be expressed as
and have N and M rules. Let x 2 X 1 and y 2 Y 1 be R1 =fx~ R1 (~x) j x~ 2 W 1 g (6)
two non-fuzzy input variables representing the states of
the robot manipulator, and z 2 Z 1 be a non-fuzzy out- where x~=(x y z ), and W 1 =X 1  Y 1  Z 1 . So, we may
put variable representing the command sent to the robot view the fuzzy relation of FN1 as a fuzzy set dened on
manipulator, where X 1  Y 1  Z 1  <. The fuzzy rules of W 1 . Similarly, the fuzzy relation R2 representing FN2
FN1 can then be expressed as can also be represented as
if X1 is A11 and Y1 is B11 then Z1 is C11  R2 =fy~ 2 R (~y) j y~ 2 W 2 g (7)
if X1 is A12 and Y1 is B12 then Z1 is C12  with y~=(x y z ), and W 2 =X 2  Y 2  Z 2 . Now, we can
dene the similarity between robot motions.
::: Denition 3 (Robot Motion Similarity) Suppose mo-
if X1 is A1N and Y1 is B1N then Z1 is C1N : tion 1 and motion 2 governed by fuzzy systems FN1 and
(2) FN2 , and let
R^i =D(Ri  n) i = 1 2:
If R^1 and R^2 satisfy the following equality: 
 
  3   
MS (R^1  R^2 )= P S  L   Q
L L

then the degree of the similarity between motion 1 and


2 is .    
#
MS ( ) is the robot motion similarity measuring opera- L
 L     Q

tor, and D( ) represents the discretelization process of


a fuzzy relation. Both of them will be described in the    

next section, and  is a real constant which belongs to 


,  $ #  #
the closed interval (0 1). L M L M

There are two kinds of methods for measuring the


similarity between two fuzzy sets: the geometric and set-  
theoretic measures 26]. Because fuzzy relations with  , J ,
similar distributions are always with dierrent coordi-
L M


nates and scales, we adopt the set-theoretic measure as P S { P S  N 
L L L L

our similarity measure. The reasons for discretelizing P S { P S  N


M M M M

fuzzy sets are convenient for computation of the set-


theoretic measure and for extracting the characteristics 
 
of distributions of the fuzzy sets. For two discrete nite P S  N  L  Q N  U 

fuzzy sets R^1 and R^2 , We can adopt a similarity measure


L L

8] to be the robot motion similarity measuring operator:


Figure 3: The robot motion classication process.
MS (R^1  R^2 )
^ ^
= jR^1 \ R^2 j (8) in the rst stage will be used to evaluate the similari-
jR1 R2 j ties between these motions using the proposed similarity
Pn3 measurement, and these robot motions can then be clas-
= Pnk3=1 min(R1 (~xk ) R2 (~yk )) (9) sied according to their similarities. In the third stage,
k=1 max(R1 (~xk ) R2 (~yk )) the CMAC-type neural network in the proposed learn-
where the operator j  j is the cardinality operator 8, ing structure will be used to learn to generalize these
11, 25]. The value evaluated using this robot motion classied motions. The processes in the three stages are
similarity measuring operator is always between 0 and shown in Figures 2, 3, and 4, respectively.
1. When R^1 and R^2 have no intersection, MS (R^1  R^ 2 ) After these three stages of learning and processing,
is equal to zero it is equal to one when R^1 is equal to fuzzy parameters will be manipulated and supplied by
R^2 . the CMAC to the FNN for the governing of input desired
Due to the variations exhibited in general robot mo- motions. The learnings in the FNN and the CMAC are in
tions, it may demand plenty of memory for a CMAC to fact interactive. When the fuzzy parameters generated
deal with dierent kinds of motions. However, if there by the CMAC are not appropriate for governing a partic-
is similarity present in a set of motions, it is easier to be ular motion, the FNN can provide more training patterns
learnt by using a CMAC of a reasonable size because of by learning to govern more sampled motions around that
their similarity. motion to increase the resolution of the CMAC. In other
words, the FNN and the CMAC will cooperate to yield
homogeneous reliable performance for a whole class of
3.3.Learning to Govern Classes of Motions motions.
By using the learning control structure described
in Sec. 3.1 and the motion similarity measurement al-
gorithm described in Sec. 3.2, classes of motions can 4.Simulation
be governed through learning. The process of learning
to govern classes of motions can be divided into three To demonstrate the eectiveness of the proposed
stages. In the rst stage, the fuzzy systems in the pro- learning control scheme, simulations of applying this sheme
posed learning structure are used to learn to govern sam- to govern motions based on using a PUMA 560 manip-
pled motions from a group of robot motions which are ulator were executed.
not yet classied. Proper parameters of the fuzzy sys-
tems which govern these motions well will then be ob-
tained. In the second stage, the parameters obtained
$Q $UELWUDU\ 5L 0RWLRQ 6/ 6 / 2


0RWLRQ 7UDMHFWRU\ 

)11V
/RFDO 9 5RERW
3ODQQHU &RQWUROOHU '\QDPLFV
2L 5L


6 6




Figure 2: Conceptual organization of robot motion governing by using a fuzzy system.

 )L  

  PL SL
S
L

    

L žQ

)
L
 

Figure 4: Conceptual organization of the CMAC-type neural network learning process.

4.1.Learning to Govern Individual Planned Mo- where Vm stands for the desired maximum joint velocity,
tions and qf the desired nal joint position. By using Eqs.(10)-
(13), we planned joint trajectories of 23 motions for the
In this stage of processing, the learning controllers PUMA manipulator. Each planned motion was denoted
were used to learn to govern a group of motions. To as mi (pi ) with i 2 f1 2 : : : 23g, and the kinematic re-
simplify the simulations, orientations were not planned lated elements of each vector pi were selected to be the
in the motions. Therefore, only the rst three joints of desired nal joint positions qf 1 qf 2 qf 3 ]t . Parameters
the PUMA manipulator were used to perform planned
i i i
of these trajectories were shown in Table 1. Because
motions. The gains of the local controllers were chosen each joint of the PUMA manipulator is equiped with an
to be xed in the simulations. Each joint of the robot FNN for governing motions, we expressed parameters
manipulator was equipped with an FNN. In each FNN, of the FNNs governing mi (pi ) as Fi = fFi1  Fi2  Fi3 g.
there are two nodes in Layer 1, fty nodes in layer 2, After performing the learning process, tracking errors
twenty ve nodes in Layer 3 and 4, respectively, and one for learning to govern each planned motions were within
node in Layer 5. A second-order system model is used to a reasonable range. The mean square errors of all the
provide critical damped reference trajectories, qr and q_r , 23 Cartesian motion trajectories at the end of the FNN
for each joint of the robot manipulator, and expressed learning were shown in Fig. 5.
as
qr (t) + bq_r (t) + kqr (t) = f (10)
with
4.2.Classifying the Governed Motions
b = 2Vm expq
(1) (11) In this stage of processing, the 23 well-governed
f motions, described in Section 4.1, were classied accord-
  ing to their similarities. Similarities between these mo-
k = Vm exp
qf
(1) 2 (12) tions are determined according to the equation below:
 = 1min 
k3 k
(14)
f = (Vm exp
qf
(1))2 (13)
= min SM(R^ik  R^jk )
1k3
(15)
  

 

Figure 5: Cartesian mean square tracking error for tracking the 23 planned motions.

 

 
  


         
   

     

     

Table 2: Motion classes with the similarity threshold of 0:2.

where R^ik and R^jk are discrete fuzzy relations calculated


from Fik and Fjk of the FNNs which govern the kth
joint of the PUMA 560 robot manipulator for motions
i and j . The number of points used for discretelization
of the fuzzy relations is 50. The similarities between
 
   
  
the 23 robot motions were then evaluated according to
     Eq. 15. To simplify our demonstration, we only classi-

    ed those governed motions according to the similarity
   
threshold value of 0:2. These motions were then divided


 
 
  




into three classes according to their degrees of similari-

    ties, and members of the three motion classes were shown

    in Table 2.
    
 
   
    
4.3.Generalizing Classied Motions to Govern
Other Similar Motions
 
   
    
  
  
    
In this stage of processing, the CMAC type neural


  
  




network was used to learn parameters of the FNNs gov-

     erning the 23 motions in three motion classes described
     in Table 2, and generalize new parameters for governing
    
other similar motions in the three motion classes.
To deal with motion class 1, the input for the CMAC
    
    
    type neural network was selected as pi dened in Sec-
   
tion 4.1. The resolution of each input dimension for
the CMAC after quantization was 0.1 rad, and the total
memory size of association cells of the CMAC was about
Table 1: Parameters of the 23 planned motions. 800k bytes. To deal with motion class 2, the input pi
for the CMAC type neural network was selected to be
the planned maxium joint velocity Vm dened in Section
4.1. The resolution of this one dimensional input vector
after quantization was 0.2 rad/sec. The size the of used
memory was 200k bytes. To deal with motion class 3, the
this, we propose a learning control scheme to enlarge
 
    
   learning space coverage by taking advantage of the mer-
      its of fuzzy systems and CMAC-type neural networks.
      
The combination of the fuzzy systems and the CMAC-
     
type neural networks in our design allows fuzzy rules
to be generalized by using CMAC-type neural networks.

  
 
    

     


This means that abstract representations are general-
ized, and they tend to cover a larger learning space.
     

Thus, the fuzzy parameters appropriate for a set of sam-


    

    

    pled motions can be generalized to deal with a whole


      class of motions. Another important factor for enlarging
     a learning space is the classication of a group of sampled
motions. This will be helpful for oerring more proper

  
 

   

   
training patterns so as to increase the speed of conver-
   
gence in the learning process and the precision rate of
   
generalization of the CMAC-type neural networks. Be-
cause of the motion classication, the generalization ca-

  
 

   

pability of the CMAC-type neural networks can be en-


   

   

    hanced in the proposed learning control scheme. Con-


sequently, motions in the same class are guaranteed to
correspond to very similar fuzzy parameters. Thus, the
Table 3: The parameters of the planned 19 motions for fuzzy parameters that the CMAC-type neural networks
generalization testing. need to generalize will exhibit much less nonlinearity and
then more classes of motions can be incorporated into the
network with a reasonable size of network memory.
input pi for the CMAC type neural network was selected Simulation results implicate that the human control
as the load carried by the PUMA manipulator. The res- strategy can be applied to robot motion control, and to
olution of this input vector after quantization was 1kg, have good performace. However, we need to know were
and the size of the used memory was also 200k bytes. about how the human makes use of the existing skills
Dimensions of the input vectors for the CMAC to deal of governing various kinds of motions to govern dierent
with motion classes are dierent, so sizes of the used kinds of motions. In our future works, we will explore
memories are dierent. useful properties of relations between dierent motion
After learning the parameters of the FNNs governing classes, develop schemes to extract motion properties of
planned motions in classes 1, 2, and 3, the CMAC type certain skills to compose another kind of skills, and study
neural networks were used to generalize new parameters what kind of format most suitable for representing mo-
for governing other 19 planned motions which belong to tion skills when used for generalization between dierent
classes 1, 2, and 3, respectively. Parameters of the 19 motion classes. All of these should be helpful for general
planned motions for generalization testing were shown robot motion governing using learning controllers.
in Table 3. The weights of all CMAC type neural net-
works were updated until errors between every training References
pattern and the corresponding reproduced parameters
were lower than 1  10;5. By using these weights, the 1] J.S. Albus, \A new approach to manipulator con-
CMAC type neural network can generalize parameters trol: the cerebellar model articulation controller
of the FNNs for governing the 19 planned test motions. (CMAC)," ASME J. Dynamic Syst., Measurement,
The mean square tracking errors for governing all the 19 Contr., vol. 97, no. 3, pp. 220{227, 1975.
Cartesian motions were shown in Fig. 6. Although the
tracking errors for generalization were large than those 2] J.S. Albus, \Data storage in the cerebellar model ar-
for learning, it was good enough for motion governing ticulation controller (CMAC)," ASME J. Dynamic
via motion generalization. Syst., Measurement, Contr., vol. 97, no. 3, pp. 228{
233, 1975.
5.Conclusion 3] C.G. Atkeson and J.M. Hollerbach, \Kinematic fea-
tures of unrestrained vertical arm movements," J.
One major problem in applying learning controllers Neurosci., vol. 5, no. 9, pp. 2318{2330, 1985.
to govern general motions of multijoint robot manipula-
tors is that the learning space is quite large. To tackle
    





    
 

Figure 6: The Cartesian mean square tracking errors for governing the 19 test motions.

4] H.R. Berenji and P. Khedkar, \Learning and tun- 14] A. Polit and E. Bizzi, \Characteristics of motor
ing fuzzy logic controllers through reinforcements," programs underlying arm movements in monkeys,"
IEEE Trans. Neural Networks, vol. 3, no. 5, pp. J. Neurophysiol., vol. 42, no. 1, pp. 183{194, 1979.
724{740, 1992.
15] T.D. Sanger, \Neural network learning control of
5] E. Bizzi, N. Accornero, W. Chapple, and N. Hogan, robot manipulators using gradually increasing task
\Posture control and trajectory formulation during diculty," IEEE Trans. Rob. and Aut., vol. 10, no.
arm movement," J. Neurosci., vol. 4, no. 11, pp. 3, pp. 323{333, 1994.
2738{2744, 1984.
16] R.A. Schmidt, Motor Control and Learning: A Be-
6] J.Q. Chen, J.H. Lu, and L.J. Chen, \Analysis havioral Emphasis, Human Kinetics, Champaign,
and synthesis of fuzzy closed-loop control systems," IL, 1988.
IEEE Trans. Syst., Man, and Cybern., vol. 25, no.
5, pp. 881{888, 1995. 17] S.J. Shiah and K.Y. Young, \Classifying robot mo-
tions to increase eciency in robot learning con-
7] J.J. Craig, Introduction to Robotics, MA: Addison- trol," in to appear in Conf. IFSA, 1999.
Wesley, Reading, 1989.
18] S.J. Shiah and K.Y. Young, \Robot motion classi-
8] D. Dubois and H. Prade, Fuzzy Sets and Sys- cation from the standpoint of learning control," to
tems: Theory and Application, Academic Press, appear in the Eighth IEEE International Conference
New York, 1980. on Fuzzy Syst., 1999.
9] T. Flash, \The control of hand equilibrium trajec- 19] T. Shibata and T. Fukuda, \Hierarchical intelligent
tories in multijoint arm movements," Biol. Cybern., control for robotic motion," IEEE Trans. Neural
vol. 57, pp. 257{274, 1987. Networks, vol. 5, pp. 823{832, 1994.
10] S.W. Keele, \Movement control in skilled motor 20] P.D. Wasserman, Neural Computing{Theory and
performance," Psych. Bull., vol. 70, pp. 387{403, Practice, Van Nostrand Reinhold, NewYork, 1989.
1968.
21] C.H. Wu, K.Y. Young, K.S. Hwang, and S. Lehman,
11] C.C. Lee, \Fuzzy logic in control systems: fuzzy \Voluntary movements for robotic control," IEEE
logic controller{Part I, II," IEEE Trans. Syst., Control Systems Magazine, vol. 12, no. 1, pp. 8{14,
Man, and Cybern., vol. 20, no. 2, pp. 404{435, 1990. 1992.
12] C.T. Lin and C.S.G. Lee, \Reinforcement struc- 22] K.Y. Young and C.C. Fan, \An approach to simplify
ture/parameter learning for neural-network-based the learning space for robot learning control," Fuzzy
fuzzy logic control systems," IEEE Trans. Fuzzy Sets and Systems, vol. 95, no. 1, pp. 23{38, 1998.
Syst., vol. 2, no. 1, pp. 46{63, 1994.
23] K.Y. Young and S.J. Shiah, \An approach to en-
13] M.C. Nechyba and Y. Xu, \Human control strategy: large learning space coverage for robot learning con-
abstraction, verication, and replication," IEEE trol," IEEE Trans. Fuzzy Syst., vol. 5, no. 4, pp.
Control Systems Magazine, vol. 17, no. 5, pp. 48{61, 511{522, 1997.
1997.
24] L.A. Zadeh, \Outline of a new approach to the
analysis of complex systems and decision process,"
IEEE Trans. Syst., Man, and Cybern., vol. 3, pp.
28{44, 1973.
25] H.J. Zimmermann, Fuzzy Set Theory and Its Ap-
plications, MA: Kluwer, Norwell, 3rd edition, 1996.
26] R. Zwick, E. Carlstein, and D.V. Budescu, \Mea-
sures of similarity among fuzzy concepts: a com-
parative analysis," Int. J. Approx. Reas., vol. 1, pp.
221{242, 1987.

You might also like