Evolutionary Electronics Companion slides for the book Bio-Inspired - - PowerPoint PPT Presentation

1

Evolutionary Electronics

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 2

Evolutionary algorithm

(schematic)

Electronic Circuit

(physical layout)

Introduction

Evolutionary Electronics (EE) is defined as the application

• f evolutionary techniques to the design (synthesis) of

electronic circuits

• Evolution of the parameters (sizing)

given a fixed circuit topology

• Evolution of both parameters and

circuit topology

• Placement and routing of the

devices that compose the circuit

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 3

Motivation: why Evolutionary Electronics ?

• Complex and performing electronic circuits are required with

which the traditional design techniques cannot cope efficiently and thus produce a waste of resources

• Some design problems are poorly

specified in the sense that the functionality is not defined in terms

• f a precise input/output

relationship but still a global performance can be measured easily

• The state of advancement of electronic technology permits the

exploitation of the virtues of evolutionary methods (waiting for the

coming of age of nanotechnologies…)

• Evolutionary electronics provides an illustration of many aspects

that are relevant in the real-world application of evolutionary methods

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 4

Electronic and Biological “Devices”

unpolarized base polarized base base collector energy emitter base collector emitter uncatalyzed catalyzed transition comp. product energy substrate

“… the fundamental elements of information processing in electronic and genetic systems are strikingly similar…”

[Simpson et al., Proc. IEEE, May 2004]

energy

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 5

Electronic and Biological Networks

“…the analysis, modeling, and simulation of natural and synthetic genetic circuits, often proceed in a manner similar to that used for electronic systems” “.. the expertise and skills contained within electrical and computer engineering disciplines apply not only to design within biological systems , but also to the development of a deeper understanding of biological functionality.” “It is possible that new strategies for engineered system design may emerge from this examination of natural gene circuit architectures.”

[Simpson et al., Proc. IEEE, May 2004]

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 6

How Evolutionary Electronics works

• 1. A genetic representation is defined for the electronic circuits
• f interest
• 2. A set of objectives is defined in terms of functionality and

performances of the desired circuit

• 3. An initial population (collection of individuals) is generated
• 4. The genome of each individual in the population is decoded

into a circuit

• 5. The functionality and performances of the circuit are

evaluated and the result is used to enforce a selection policy

• 6. The selected individuals are reproduced and the genetic
• perators are applied to obtain a new population

Until the objectives are met or some stopping criterion (time, computational resources…) is fulfilled

7

Extrinsic vs. Intrinsic evolution

The behavior of an electronic circuit can be estimated using a circuit simulator In extrinsic evolutionary design: each circuit is simulated to assess its performance

* SPICE simulation example * circuit description: C1 2 3 10n C2 4 5 100n R1 3 1 270k R2 4 1 1k R3 0 5 1k Q1 4 3 0 Q2N2222 V1 1 0 5Vdc * input signal: Vin 2 0 SIN 0 0.2 1k 0 0 0 * simulator directives: .TRAN 0 3.5ms 0 0.01ms .OPTIONS TEMP=25 * device models: .MODEL Q2N2222 NPN + IS 14.3E-15 BF 256 NF 1 + VAF 74 IKF 0.28 ISE 14.3E-15 + NE 1.3 BR 6.1 NR 1 RB 10 RC 1 + CJE 22.0E-12 MJE 0.377 + CJC 7.3E-12 MJC 0.3416 + TF 411E-12 XTF 3 VTF 1.7 + ITF 0.6 TR 46.9E-09 XT B 1.5 .END

time (ms)

3.0 2.0 3.5 0.0 1.0 0.8 0.6 0.4 0.2

• 0.2
• 0.4
• 0.6
• 0.8
• 1.0

voltage (V)

Vin Vout

0.5 2.5 1.0 1.5

• The structure of the circuit can be almost

arbitrary

• The circuit is virtual and cannot be damaged
• The simulated models are only approximations

C2 100nF Q1 2N2222 V1 5V Vin Vout R2 1kΩ R1 270kΩ

2 3 1 4 5

R 1 3 kΩ C1 10nF

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 8

In intrinsic evolutionary design: each circuit is physically implemented and tested The existence of reconfigurable devices opens the way to the possibility of performing evolution directly in hardware

... ... ... ...

cell input/output circuitry connections

• The evaluation is done using real

devices (no approximations)

• There are some constraints on the circuit

structure

• The results can depend on the physical

device used for the evolution

• The evolutionary setting can be defined so as to produce an

adaptation during the operating life of the circuit (e.g., to obtain fault tolerance)

Extrinsic vs. Intrinsic evolution

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 9

Analog vs. Digital circuit evolution

Analog signals

time amplitude clock time amplitude

Digital signals

C2 100nF Q1 2N2222 V1 5V Vin Vout R2 1kΩ R1 270kΩ

2 3 1 4 5

R 1 3 kΩ C1 10nF

Analog circuit Digital circuit

• Functionality determined mostly by

connectivity; few circuit parameters

• Discontinuous search space
• Functionality determined by

connectivity and circuit parameters

• Mostly continuous search space

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 10

Conventional digital design

• Basic building blocks at various levels (silicon, transistor, gate,

functional)

• Combinational and sequential circuits
• Objectives specified in terms of truth tables, state machines,

hardware description languages (HDL…)

• There exist systematic techniques for

combinational and sequential design which typically combine building blocks

• There is a large variety of commercial

reconfigurable (programmable) devices

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 11

Conventional analog design

• Basic building blocks at various levels (silicon, transistor, functional)
• There are fewer analog than digital

high-level standard building blocks

• Multiple objectives specified in terms
• f I/O transfer functions, power

consumption, noise, operating voltage ranges…

• No systematic methodologies for

analog design. The creativity, intuition, experience of the designer is required to obtain state-of-the-art

• performance. Analog design is still an

art rather than a science

• There are few commercial analog

reconfigurable devices

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 12

• Abstraction is the process of disregarding some aspect of a

system focusing on a subset of its actual properties

• Abstraction is important in human design since it reduces the

number of elements that must be taken into account simultaneously

• The drawback (“cost of abstraction” or “abstraction-efficiency

tradeoff”) is the less efficient use of the available resources since some system configurations are no longer considered

• Artificial evolution can proceed with different abstractions and

explore parts of the design space that are not accessible to a human designer

The role of abstraction

Space of all electronic circuits Current space of evolutionary designs Current space of conventional designs

• New design principles

can be discovered by analyzing the workings

• f evolved circuits

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 13

The Story of the P2

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 14

Evolutionary digital design

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 15

At a glance…

• Digital evolutionary design is mostly combinational
• Circuits are typically encoded as binary strings. The string is the

chromosome used by evolutionary algorithms

• The most common genetic representations are
• Schematic-based: the genome contains a description of a

schematic, for example the functionality of circuit elements and the interconnections between those elements

=

• Truth table-based: the content of the

truth table is evolved. The genotype contains the outputs of the circuit for each possible input.

• Configuration bits-based: the evolution is performed on a

reconfigurable component whose functionality is defined by configuration bits. The genome is composed directly by those configuration bits

• Fitness typically depends on the number of correct I/O relations
• r on the definition of a global measure of performance

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 16

Reconfigurable digital devices I

Programmable components that can implement a user specified digital circuits. Programmability is achieved with configuration bits that are downloaded to the component and which represent how logic gates are wired up inside of it.

Out0 = In0 · In1 · … + + In0 · In1 · … + … Out1 = … Out2 = …

Programmable Array Logic (PAL)

• Used to compute sum-of-product terms
• Permits the specification of the connection between the inputs

and the product lines

Out0 Out1 In0 In1 (Note for the laboratory: obtaining outputs fixed at 0 or 1)

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 17

Reconfigurable digital devices II

Programmable Logic Array (PLA)

• used to compute sum-of-product terms
• permits the specification of the connection between the inputs and the

product lines and between the product outputs and the sum lines PAL PLA Custom reconfigurable digital devices

• Many kinds (e.g. POEtic Circuit, especially designed for evolutionary

applications)

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 18

Reconfigurable digital devices III

Field Programmable Gate Array (FPG A)

• Device composed of an array of programmable functional blocks (cells)
• Typically each cell can generate arbitrary combinational functions of a few

number of inputs and includes some memory elements

• Includes a large reconfigurable network of interconnections between the cells
• The functionality of the cells and the interconnections are described by

configuration bits whose configuration is downloaded into the chip

Example: the XC 6200 family of FPGAs from Xilinx, which is often used for EE applications since the specification of the configuration bits is available and no configuration can damage the chip (this permits unconstrained evolution)

MUX

Func. Unit

MUX MUX

N S E W N S E W N S E W

MUX

S E W F N

MUX

N E W F S

MUX

N S E F W

MUX

N S W F E

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 19

• The goal is the evolution of a

controller for robot navigation

• The robot is put in a square arena

with obstacles that do not allow free navigation but let the robot see the target (X)

Example: Evolution of a robot controller

[Keymeulen et al., 1998]

1 2 49 50 1 2 3 1 2 8 motor decoder Sensors encoder

PLA

Obstacle detectors detector

• f

direction

M M

...

• A PLA connects the obstacle and target

detectors to a decoder that drives the wheel motors

• The configuration bits of the PLA

constitute the genome

• The fitness represents the progress

from 64 predefined starting positions to the target

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 20

Genetic representation: – The outputs of the functional cells are sequentially numbered – For each cell a few bits encode the connections of the inputs and the block function – Feedback loops are avoided (evolution is constrained)

• Among the results, standard circuits such as

adders with unconventional topologies or less gates than the smaller hand-designed circuit for the same function

Example: Evolution of arithmetic circuits:

Cartesian Genetic Programming [Miller et al., 2000]

Conventional two-bit multiplier Evolved two-bit multiplier

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 21

Example: Evolution of arithmetic circuits II

[Miller et al., 2000] Conventional three-bit multiplier (24 two-input gates + two multiplexers) Evolved three-bit multiplier (24 two-input gates)

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 22

• The goal is to discriminate between two frequencies (1KHz/10KHz) using 100 cells of a

Xilinx XC 6200 FPGA

Example: unconstrained evolution

[Thompson et al., 1999]

• The genome is constituted

by the 1800 configuration bits of the cells, which specify the connectivity and functionality of the cell

• The input and output lines

are predefined

• No external clock is provided

to the cells

• No constraints are imposed

to the topology (feedback loops allowed)

• The circuits are tested on

bursts of the two kinds of signals

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 23 Generations Evolved circuit Functional part

• The evolved circuit is very compact
• It uses internal dynamics (feedback loops), physical characteristics (gate propagation time),

and electromagnetic interactions not contemplated by the conventional design approach

• The experimenters could not understand how the circuit achieves its functionality
• The solution is sensitive to environmental conditions and to changes of the physical device

Unconstrained evolution: Results

[Thompson et al., 1999]

Functional part as digital circuit

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 24

Evolutionary analog design

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 25

At a glance…

• Analog circuit evolution imposes less constraints, allows more interactions and

has more evolutionary potential than the evolution of digital circuits

• The absence of systematic analog design methodologies increases the interest

in automatic synthesis solutions

• Analog design is intrinsically multi-objective (multi-objective evolutionary

algorithms)

• Some possible genetic representations are
• Configuration bits-based: the genome is composed directly by those

configuration bits of the reconfigurable component

• Schematic-based: the genome contains a description of a schematic, for

example the functionality of circuit elements and the interconnections between those elements

• Tree-based: the genome is a tree structure which defines a construction

process for the circuit, starting from an elementary circuit (“embryo” + test fixture)

• Sequence matching-based: the interaction between the devices is

implicitly defined by the genome in terms of sequence matching

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 26

Reconfigurable analog devices I

Field Programmable Analog Array (FPAA)

• Device composed of an array of programmable

functional blocks (cells).

• Typically each cell is built around an operational

amplifier

• It can be used to implement filters, oscillators,

amplifiers, comparators…

• The functionality of the cells and the

interconnections are described by configuration bits whose configuration is downloaded into the chip

FPAA array of configurable analog blocks (CAB) Detail of the analog block

Example: The AN10E40 FPAA from

• Anadigm. It is composed by a 5x4 array
• f blocks containing an operational

amplifier with programmable interconnections and on-chip programmable resistors and capacitors.

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 27

Reconfigurable analog devices II

Evolvable Motherboard (EM)

• The EM is an array of programmable

analog switches

• External discrete components are

connected to the matrix with daughterboards.

• The genetic representation corresponds

to the configuration of the switches (1 bit per switch)

• The components can be of any type: low-level components (e.g. transistors) for

fine-grained evolution, or high-level components (e.g. amplifiers, logic gates) for coarse grained evolution.

Example: Evolution

• f an inverter

with bipolar transistors Evolved circuit

generations Fitness

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 28

Schematic-based representation

• The experimenter defines

the kinds of electronic devices that can appear in the evolved circuit

• Each “gene” represents a

device and specifies its parameters and connectivity

• All the components and

connections must be explicitly represented in the genome

• The genetic representation

can produce dangling components that must be removed before simulation

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 29

Tree-based representation (Genetic Programming - GP)

• The experimenter defines the

kinds of electronic devices that can appear in the evolved circuit

• The experimenter defines set
• f circuit-modifying
• perations that alter the

topology and parameters of the circuit

• The genome represents a

program of circuit development

• The development starts from

an “embryo” that contains the fixed sources and loads for the evolved circuit

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 30

Sequence matching-based representation

(Analog Genetic Encoding – AGE)

• The experimenter defines the kinds
• f electronic devices that can appear

in the evolved circuit

• Each “gene” represents a device and

specifies sequences of characters associated with its terminals and parameters

• The interactions (determined by the

resistors) are implicitly defined by the device interaction map f, rather than explicitly represented in the genome

• Sources and loads are assigned as a

fixed circuit

• The genetic representation can

produce dangling components that must be removed before simulation

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 31

Example: evolution of a Gaussian function generator

• The goal is the evolution of a circuit that produces an output current that is

a Gaussian function of a variable input voltage

Evolved circuit I/O behavior

Sequence matching-based (AGE) Tree-based (GP)

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 32

Further issues

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 33

Design Verification

• A designed circuit must be robust and perform correctly within a given
• perational envelope defined by the combined range of all the expected
• perational conditions and parameter values (power supply voltages,

environment temperature, device variability and aging…)

• Conventional design comes equipped with verification techniques
• The evolutionary process
• Tends to take into account only the conditions in which the fitness is

evaluated

• Can produce non conventional circuits that are difficult to understand

and to verify with conventional techniques Therefore, in the use of evolutionary electronics

• The definition of the fitness must involve and extensive sampling of the
• perational envelope
• The evolved circuit must be tested to assess its generalization properties
• The possibility of including online adaptive potential must be considered

Companion slides for the book Bio-Inspired Artificial Intelligence: Theories, Methods, and Technologies by Dario Floreano and Claudio Mattiussi, MIT Press 34

• Evolution can be performed in simulation or on the actual physical system
• The synthesis of an system can be done in the absence of a complete

specification and with minimal human intervention in the design process

• The synthesis via evolution can achieve better use of resource relatively to

a conventional design process

• The evolutionary process can discover new system configurations
• Evolutionary electronics can be applied in contexts where systematic

methodologies of conventional design do not exist

• Real-world evolutionary synthesis typically has multiple objectives
• We must keep in mind the important issue of design verification

Summing up…

Evolutionary electronics illustrates many issues of the real-world application of artificial evolutionary techniques, for example