Computational Synthesis of Multi-Domain Systems
Proceedings of the 2003 AAAI Spring Symposium -
Computational Synthesis: From Basic Building Blocks to
High Level Functionality, page 59--66. Stanford, California, (March 2003)Abstract
Several challenging issues have to be addressed for
automated synthesis of multi-domain systems. First,
design of interdisciplinary (multi-domain) engineering
systems, such as mechatronic systems, differs from
design of single-domain systems, such as electronic
circuits, mechanisms, and fluid power systems, in part
because of the need to integrate the several distinct
domain characteristics in predicting system behavior.
Second, a mechanism is needed to automatically select
useful elements from the building block repertoire,
construct them into a system, evaluate the system and
then reconfigure the system structure to achieve better
performance. Dynamic system models based on diverse
branches of engineering science can be expressed using
the notation of bond graphs, based on energy and
information flow. One may construct models of
electrical, mechanical, magnetic, hydraulic, pneumatic,
thermal, and other systems using only a rather small
set of ideal elements as building blocks. Another
useful tool, genetic programming, is a powerful method
for creating and evolving novel design structures in an
open-ended manner. Through definition of a set of
constructor functions, a genotype tree is created for
each individual in each generation. The process of
evaluating the genotype tree maps the genotype into a
phenotype -- i.e., to the abstract topological
description of the design of a multi-domain system,
using a bond graph along with parameters for each
component, if needed. Finally, physical realization is
carried out to relate each abstract element of the bond
graph to corresponding components in various physical
domains. To implement the above GPBG approach in a
specific application domain, cautious steps have to be
taken to make the evolved design represented by bond
graphs realizable and manufacturable. To achieve this,
one important step is to define appropriate building
blocks of the design space and carefully design a
realizable function set in genetic programming. We are
going to illustrate this in an example of behavioral
synthesis of an RF MEM circuit C a micro-mechanical
band pass filter design. Finally, we have some
discussions on how to extend the above approach to an
integrated evolutionary synthesis environment for MEMS
across a variety of design layers.