Designer-driven Stochastic Multi-objective Optimization for
Circuit Sizing and Hierarchical synthesis
A New Methodology for analog circuit
optimization
We have designed a new methodology for automatic sizing of analog/RF circuits by incorporating
the designer's knowledge/intuition into simulation-based stochastic
circuit optimization. Our technique uses SPICE in loop for optimization, however
is not 'blind' or 'random' in search like earlier approaches by Rutenbar, et.al.
and Gielen, et.al. We have a methodology to capture designer's intuition in a
simple table with choices (also derivable from simple first order equations) and
use this information to guide the adaptive search leading to fast and accurate
optimization.
Features
- Method to captures designer's intuition in a simple table with choices.
- Tabular information also derivable from simple first order equations
- More accurate optimization without any additional computational expense
- Around 10 times faster optimization than 0-knowledge stochastic
optimization
- No need of SPICE accurate equations, needs only approximate equations
- Optimization is SPICE accurate
- Very easy to fill in the required information for designer and scalable
to large circuits.
What we are not
- We are not equation-based systems (like AMGIE, 2001) or hybrid equation
and simulation-based systems (like ASTRX/OBLX, 1994) . These approaches need
accurate equations and use them to replace simulations for evaluation. We
don't need accurate equations and we don't use equations for evaluations.
- We are not 'interactive optimization', where the designer biases the
selection process of the evolutionary algorithm. We don't bias selection
process by designers' choice, though that can be added as a feature to the
system. We capture the designer's intuition once in an innovative
representation and use it to improve the optimization.
- Our approach is not one algorithm. It is a technology which can be used
to improve any black-box optimization algorithm for circuits, both with
multi-objective and single-objective formulations. We show the application
of the technology to NSGA-II, but it is not limited to it.
- We are not geometric programming and do not require SPICE-correct
equations.
Limitation of other approaches transcended
in current methodology
Geometric Programming
- Needs SPICE accurate circuit equations
- Equations constrained to posynomial form. Does not generalize to all
circuits.
- SPICE-correctness of optimization not assured.
- Lot of manual work on part of designer to setup accurate equations
Simulation Based Approach
- Blind and not circuit-specific: Slow and inaccurate
- Does not exploit designer's intuition about the circuit
Technique Description
- 2-slide summary of methodology flow and results [pdf,
img]
- V. Aggarwal, U.M. O'Reilly, "Structural information in simulation-based
approaches for efficient circuit sizing", (Paper under Submission)
- V. Aggarwal, U.M. O'Reilly, Paper on multi-objective optimization theory
and circuit optimization, Under Submission to GECCO 2007
Other information
- This work is under MIT Patent Filing.
- Authors:
- Varun
Aggarwal, PhD Student, Computer Science and Artificial Intelligence
Laboratory, MIT, USA; Contact: varun_ag [at] mit DOT edu
-
Una-May O'Reilly, Principal Research Scientist, Computer Science and
Artificial Intelligence Laboratory, MIT, USA; Contact unamay [at] csail
DOT mit DOT edu
- Group: EvoDesignOpt,
CSAIL, MIT, USA
Other/Forthcoming Reports
- Piecing the puzzle: Equation-based approaches, Simulation-based
approaches and Geometric Programming
- How to do scalable circuit optimization through reuse.
- Circuit-Open-Encyclopedia: The concept, the architecture and the future.
- Handling dynamic linear constraints for circuit optimization
- DC Operating Point Formulation for circuit optimization: Why and how?
For more information, contact
Varun Aggarwal, varun_ag [at] mit DOT edu