A computational theory for control of posture and movement in a multi-joint limb. - Page 198 |
Save page Remove page | Previous | 198 of 252 | Next |
|
small (250x250 max)
medium (500x500 max)
Large (1000x1000 max)
Extra Large
large ( > 500x500)
Full Resolution
All (PDF)
|
This page
All
|
1 8 3
however, its main drawback was that since it was conceived to function in repetitive
tasks, it could not generalize, even to follow very similar trajectories.
Recent applications of neural networks in implementing some of these adaptive
control schemes have provided examples of how it is possible to build reasonably
accurate internal models from repeated movements, i.e., through practice. Many of the
current models which use a distributed form of representation (e.g., a look-up table) have
a learning scheme that is based on the idea that experience in one activity should improve
performance in all similar activities (Kohonen 1982, Miller et al. 1987, Ritter and
Schulten 1988). This notion of local generalization takes advantage of the fact that the
inverse dynamic transformation is continuous, in that similar inputs (desired trajectories)
lead to similar outputs (torques or muscle activations). Tabular form is well suited to this
kind of transformation because table entries can be shared between similar input patterns,
and interpolation can be made to smooth the transformation between neighboring inputs.
A more powerful kind of representation is one that captures the inherent structure
of the transformation, i.e., global generalization. For the equations of inverse dynamics,
Kawato et al. (1987) and Miyamoto et al. (1988) implemented a system where the
adaptive controller had to learn the coefficients of a set of non-linear units, where the
output of the units were weighted by these coefficients and summed to form a joint
torque, and each unit estimated a part of the inverse dynamics equation. This kind of
representation is critically dependent on the internal model of the controlled arm (e.g.,
content of the units), but in terms of learning efficiency, because global generalization
incorporates knowledge of the structure of the motor apparatus, it typically improves
more quickly than table based models for the same amount of experience.
Consider the problem of lifting a load, where the dynamics of the arm may change
dramatically as compared to when the load is only the arm itself. Structured models can
be updated quickly since only a few parameters need to be updated, and this learning
Object Description
Description
| Title | A computational theory for control of posture and movement in a multi-joint limb. - Page 198 |
| Repository email | cisadmin@lib.usc.edu |
| Full text | 1 8 3 however, its main drawback was that since it was conceived to function in repetitive tasks, it could not generalize, even to follow very similar trajectories. Recent applications of neural networks in implementing some of these adaptive control schemes have provided examples of how it is possible to build reasonably accurate internal models from repeated movements, i.e., through practice. Many of the current models which use a distributed form of representation (e.g., a look-up table) have a learning scheme that is based on the idea that experience in one activity should improve performance in all similar activities (Kohonen 1982, Miller et al. 1987, Ritter and Schulten 1988). This notion of local generalization takes advantage of the fact that the inverse dynamic transformation is continuous, in that similar inputs (desired trajectories) lead to similar outputs (torques or muscle activations). Tabular form is well suited to this kind of transformation because table entries can be shared between similar input patterns, and interpolation can be made to smooth the transformation between neighboring inputs. A more powerful kind of representation is one that captures the inherent structure of the transformation, i.e., global generalization. For the equations of inverse dynamics, Kawato et al. (1987) and Miyamoto et al. (1988) implemented a system where the adaptive controller had to learn the coefficients of a set of non-linear units, where the output of the units were weighted by these coefficients and summed to form a joint torque, and each unit estimated a part of the inverse dynamics equation. This kind of representation is critically dependent on the internal model of the controlled arm (e.g., content of the units), but in terms of learning efficiency, because global generalization incorporates knowledge of the structure of the motor apparatus, it typically improves more quickly than table based models for the same amount of experience. Consider the problem of lifting a load, where the dynamics of the arm may change dramatically as compared to when the load is only the arm itself. Structured models can be updated quickly since only a few parameters need to be updated, and this learning |
