Stochastic Optimization with Learning For Complex Problems
Stochastic Optimization with Learning for Complex Problems
Original Proposal: Stochastic Optimizer with Learning
Motivation and Scope
High-Level Goals
Data Engineering
Modification to Simulated Annealing
Selecting a Candidate Move
Three Approaches
Parameters in the Model
Cost Estimate vs. TimberWolf Chip Area
Overall Flow of Experiment
Performance of Trained Model: Apex6
Variation of Parameter Model Weights with Temperature Zone
Determination of Learning Sample Size
How Many Swaps-per-Temperature? (Training Set)
How Many Swaps per Temperature?
Discovering the “Windowing Effect”
Windowing Effect
Normalized Swap Distance
Training Time
Runtime Performance
Handling Multiple Conflicting Objectives
Multiple Objective Functions: Trading Off
Example Tradeoffs: C5315
Netlist Partitioning
Examples for Partitioning
Dutt and Deng’s PROP
Decision-Theoretic FM
BDD Sifting Evaluation: Default vs “Optimal”
BDD Sifting Evaluation
Potential Improvement for BDD Optimization
Summary
Email: rnewton@ic.eecs.berkeley.edu
Home Page: http://www-cad.eecs.berkeley.edu/~newton