Feature Selection for Instruction Placement for the TRIPS Microarchitecture
Katie Coons and Bert Maher
Project Documents
Project
Proposal
Meeting
Notes - March 8th, 2007
Mid-semester
Writeup
Final
Report, and the final perfomrance
results, which we were not able to include in the paper, as condor
was being fussy about running our jobs. All benchmarks/parameters are
identical for these two experiments, except that in one case we used the
feature set currently in the scheduler and in the other case we used a
modified set of features based on the lasso/correlation results.
The Y axis in this graph is the fitness of the best organism. Fitness is
cacluated as percent improvement over the scheduler + 2, because all
fitnesses in NEAT must be positive. Thus, the line where the y-axis
equals 2 is where NEAT is performing exactly as well as the sheduler -
the percent difference is zero. The best results we've seen so far
for this benchmark set are less than a percent worse for the initial
features, and around 3% better than the scheduler for the lasso features.
This is for a benchmark set that did particularly poorly with NEAT.
We're trying the new features out on a benchmark set that did notably
better with NEAT now (5% better than the scheduler). Hopefully we can
improve that result as well.
Matlab code, data, and other source files for the final project can be
found here.
Machine Learning for Instruction Placement for the TRIPS Microarchitecture
Katie Coons, Behnam Robatmili, and Matt Taylor
Project Documents
Project Proposal
Interface Specification
Final
Report
References
A
spatial path scheduling algorithm for EDGE architectures
Merging
head and tail duplication for convergent hyperblock formation
Building a
basic block instruction scheduler with reinforcement learning and
rollouts
Evolving neural networks
through augmenting topologies
Meta
optimization: improving compiler heuristics with machine learning.
Comparing
evolutionary and temporal difference methods in a reinforcement learning
domain