Hardware Reference
In-Depth Information
4.5
General Validation Flow of RSMs
To verify the quality of the predictions generated by a RSM we will compare the
predicted metric values with the actually observed ones to obtain then a quality index
for the model. To do so, we introduce the average normalized error :
x
ρ ( x ) i
y ( x ) i
y ( x ) i
i
|
|
η
=
(4.20)
|
|
where:
￿
is the set of system metrics,
￿
is the set containing all the design space points,
￿
y ( x ) i is the actual value of the metric i
for the design space point x
, and
￿
ρ ( x ) i is the estimated value of metric i
for the design space point x
.
An appropriate RSM should present a behavior where the greater is the training set
size, the better is the average normalized error. As a matter of fact, we will verify that
the error decreases growing the training set size, and this is done re-running the model
construction with a greater training sets and calculating the average normalized error.
If random processes are involved during the selection of the design space points
used as training set, the prediction results can present variability (especially for
training sets with small size with respect to the design space). To characterize this
effect, the validation methodology will be repeated to identify a set of stable statistical
properties.
4.6
Conclusions
In this chapter, we have introduced a set of statistical and machine learning tech-
niques that can be used to improve the performance and/or accuracy of design space
exploration techniques by predicting system level metrics without resorting to long
simulations. The presented techniques (called Response Surface Models) leverage
a set of closed-form analytical expressions to infer an approximation of the actual
system response by either exploiting regression or interpolation modeling. Results
of the validation of RSM algorithms will be described in Chap. 8.
References
1. Fillon, C.: New strategies for efficient and practical genetic programming.
Ph.D. thesis,
Università degli Studi di Trieste (2008)
 
Search WWH ::




Custom Search