broken sensor? no problem! imagine being able to continue using a physically damaged sensor – how...

1
BROKEN SENSOR? NO PROBLEM! Imagine being able to continue using a physically damaged sensor – how could you do it? Imagine being able to extend the range of a sensor 2 or 3 times, with no extra cost to your supplier? Imagine sensor processing tasks that are no longer limited to strict environmental conditions, sensor capabilities, or even to particular sensors… With dynamically reconfigurable analogue circuits these scenarios are becoming a possibility. Of course the technical difficulty of designing analogue circuits hasn’t gone away, but we now have the software technology to allow us to discover analogue solutions that might take years for human engineers to discover. Each chip interfacing with a sensor can be reconfigured as many times as needed to ensure a linear response is met as conditions change. New circuits can be evolved in situ. Combinations of sensor inputs can be controlled and analysed by combinations of chips in circuit configurations that would have conventional analogue engineers in dismay! Evolution allows this because it is a “test it and see” approach. No calculation is necessary. No engineering is carried out. OUR INSPIRATION It can be a shock to discover how often nature re-uses genes that make good solutions. Genes responsible for bone material, the formation of limbs, sight and a host of other morphological features are widely re-used in different contexts across species. The context in which a gene actives the development of morphological feature can affect its size (such as the decreasing size of vertebra in your spine), shape (such as appendages) or even function. But what is the mechanism for gene re-use? How is that certain genes are expressed in particular locations in the embryo and not others? Developmental processes create repeated morphological structures due to the distribution of binding proteins throughout the embryo. The presence of proteins that bind onto sections of the DNA acts like a “switch” that determines which genes are expressed in which location. Thus DNA and its associated RNA transcription processes produce proteins that affect the production of other proteins later in the developmental process. In effect, DNA produces the rules that govern its own production of other rules and other proteins! A similar mechanism is used to control the body’s responses to changes in the environment. The presence of a particular protein in a cell (such as an auxin inhibiter) may trigger a chain of other proteins being produced that lead to the cell undergoing mitosis. BACKGROUND Evolutionary computation has been hampered by an inability to tackle problems of significant complexity or scale. Although it has achieved notable successes, its application range has extended little beyond search-based optimisation tasks. To date, no one has evolved a car, a house or anything where the number of parts exposed to mutation and selection is large and interconnected. However, Thompson’s (1996) ground-breaking hardware experiments showed that evolution can be used to discover solutions in parts of the design space normally inaccessible to human engineers. Evolution is able to operate in design domains whose complexity and physical richness would defeat human engineers, who usually try to simplify or abstract away from physical reality in order to make their problems tractable. RICH PHYSICAL SEARCH SPACES The use of evolutionary algorithms to discover novel (even patentable) analogue circuits has a relatively long academic history. The evolvable hardware initiative from NASA also explored the innate complexity of physical search spaces. However, for the first time, we have the opportunity to explore a physical medium that is capable of continuous real-time reconfiguration. The new dynamically programmable Analogue Signal Processors (dpASPs) are destined to replace many ASICs (application specific integrated circuits), and offer potentially disruptive technology wherever digital logic has to interface with the real world. These chips offer us the chance to return to the complex world of analogue signal processing with new eyes: eyes provided for us by evolutionary search algorithms. Binding Analogue Signals To Digital Genomes: using feedback to guide evolutionary search HOX GENE EXPRESSION DURING DROSOPHILA DEVELOPMENT Stripes, spots and dashes: the presence of binding proteins distributed in varying degrees of concentration in individual cells allows gene “switches” to be activated. Which genes are expressed (or inhibited) in turn leads to the context specific production of proteins at these locations, giving rise to repeated, modular morphological structures such as body segments, wing disks, etc. Credit: Dave Kosman, UCSD Kester Clegg ( [email protected] ) Supervisor: Prof. Susan Stepney ( [email protected] ) http://www.cs.york.ac.uk/~kester PATTERN FORMATION AND MODULAR RE-USE Top: The presence of different binding proteins combine to produce spots on a drosophila embryo. Bottom: Different switches causing the gene for Bone Material Protein 5 to be expressed in different locations in a mouse embryo. The resulting morphology is context dependent. Image taken from Carroll (2006:125) THE TECHNOLOGY The Jacyl AXR-16 board: Contains 4 Anadigm FPAA chips (shown on left) which can be daisy-chained together. The chips use switched capacitors to reproduce analogue components. The reconfiguration of the chips can be controlled using the on-board FPGA (larger chip on right) or by an externally attached PC. Each of the FPAA chips can accept up to 4 inputs and have up to two outputs. Their reconfiguration can be independent of one another. An API for the reconfiguration software is provided by Andadigm. AnadigmDesigner2 Anadigm provide software to be used as a GUI design tool for testing and implementing analogue circuits on their chips. However, the software also has an API which can be ‘wrapped’ in other code, thereby automating the design process and downloading of circuits. EXTERNAL INPUT OSCILLOSCOPE PROBES Oscilloscope probes read the analogue signals output by the dpASP chips. The signal is passed through an ADC and then matched against the genome to see if the “signal” can bind to a part of the genome. If it can, then the circuit is re-configured according to the genome’s specification. GENERATION OF THE PHENOTYPE: PRUNING DANGLING NODES A visualisation of the graphs show that randomly generated wiring is haphazard, improbable, and often unusable! The top graph of 50 nodes and 8 dangling outputs needs to be pruned until a single dangling output is left (if one chip is used) as shown in the bottom graph. Even very large graphs of a 100 nodes and more collapse quickly when pruned, meaning that most of the genotype is not realised in the phenotype. However, this “junk” is still subject to mutation and may come into use in a later generation if the mutation is beneficial. STATE-DRIVEN DYNAMIC RECONFIGURATION State-driven dynamic reconfiguration works by continuously analysing the state of a chip. If a particular state is recognised, then the reconfiguration process is triggered. The Anadigm software only uploads the differences between the last circuit and the previous circuit, so reconfiguration is as fast as possible. A single CAM can have its parameters tweaked, or an entirely new circuit can be uploaded. In this picture, changing input from the sensors triggers a reconfiguration of the circuit as needed. METHOD AND PLATFORM Our “evo-devo” platform consists of 4 FPAA chips daisy-chained together. Each of the chips can receive multiple inputs and each can be re-configured with a new analogue circuit independently. The reconfigurations can be controlled by the on-board FPGA chip or from a PC connected to the board. The genomes are first developed using the software shown on the left. Each genome is then “pruned” until the required numbers of dangling outputs are left. Dangling outputs can be fed into the next chip or to an external source. The phenotypes are then translated into circuit specifications. Circuits which are unable to fit on a chip (due to resource considerations) are discarded. However, circuits which have incompatible configurations (due to meaningless connections, different clock speeds, bizarre feedback ‘islands’ etc.) are only discarded if the Anadigm software prevents us from downloading the circuit. We are not too concerned by warnings from the software that these circuits cannot be analysed! The mechanism of triggering the reconfiguration is unique, as it attempts to replicate the feedback process that DNA and binding proteins achieve to control further protein production. By hooking oscilloscope probes into the circuit, we hope to match the analogue signals with potential “binding sites” on the digitally represented genome. For this to be possible, the oscilloscope software passes analogue signals through an ADC and the signal values are fitted (via fast Fourier transform) against the bin values schema shown on the right. Unlike most evolutionary computation, our genome is never fully realised at any one time. Instead, a large part of the genome contains “switches” that may or may not be activated on receiving a matching signal, leading to potential reconfigurations lying “dormant”, waiting for the right circuit conditions to appear. These switches are still exposed to the mutation of evolutionary processes and may be judged as part of the DNA “junk” if they are not used. GENERATION OF GENOMES Left panel: shows details of CAMs (Configurable Analogue Modules). Each CAM can be configured with different parameters, clock speeds and option settings to give different signal processing capabilities. These configurable modules form around 168 building blocks that are represented in the genome shown on the right. Right panel: using Cartesian Genetic Programming (CGP – developed by Julian Miller, now at York), a feed-forward directed graph is generated with each node in a graph representing a CAM configuration. Feed-forward means that all inputs to nodes are guaranteed to be taken, but outputs can be left dangling and during the phenotype generation these need to be pruned. NEUTRAL VARIATION Evolution is remarkable as a search algorithm. Its ability to hill-climb is well known and has been used in many optimisation applications. However, it is as an algorithm of invention that it really stands out. When combined with rich physical environments, evolutionary algorithms frequently find solutions that are judged incomprehensible to human engineers. Adrian Thompson (1998) has suggested this is in part due to neutral variation in the genome allowing the algorithm to “jump” from peak to peak in the search landscape, as the neutral part of the genome is never realised or selected for. Evolution appears to be the only search algorithm capable of reaching solutions in this way. THE WILDCARD BINDING SCHEME We can only be inspired at how binding proteins attach to DNA and the complex chemistry involved. The mechanism of self-regulation and context-specific response to environmental change is fundamental to how organisms are controlled. In our model, the binding proteins become analogue signals with their matching binding sites on the genome taking a digital representation. In wildcard positions, any value for the converted signal will match , making the binding sites flexible in terms of what range of signals they will accept and how they will act as switches. Analogue signal activity therefore has the potential to feedback into the genome, triggering a reconfiguration when certain conditions are met. HOW DO PROTEINS BIND TO DNA? Proteins attach to stretches of DNA material by recognising signature sequences of base pairs. For example, a single switch for a gene may consist of several hundred base pairs (bp), lying perhaps several thousand bp upstream of the gene. Within the gene switch, there are usually 6-20 signature sequences (each 6-9 bp in length) that affect the expression of the gene concerned. Signatures are therefore very small, but this does not mean they are limited in their potential combinations (in fact, even a relatively short signature has a huge number of combinations). Signature sequences are sometimes exact for every position, sometimes they contain wildcards. Wildcard positions can be filled by all four nucleic acids (Cytosine, Thymine, Adenosine, and Guanine) but are more often limited to pairs of alternatives (e.g T or A, C or G, etc.). For example, Tinman, a gene related to heart development in most species, from drosophila to humans, is highly specific (this example is taken from Carroll (2006)): Tinman TCAAGTG Pax-6 (eyeless) KKYMCGCWTSATKMNY Dorsal GGGWWWCCM whereas Pax-6 (the master gene controlling forms of sight in species), and Dorsal, use the wildcards represented by K (G or T), Y (C or T), M (C or A), W, etc. Thus Pax-6 has a signature with only 6 specific sites out of 16 possible bp combinations, meaning that it could bind at a variety of locations. Our analogue signal binding scheme is shown below. It too has wildcards as bin values representing the transformed analogue signals. Like nature, we allow evolution to mutate these binding sites, allowing new “switches” to evolve that have different binding properties, allowing new genes to be switched on.

Upload: anna-carlson

Post on 27-Mar-2015

217 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: BROKEN SENSOR? NO PROBLEM! Imagine being able to continue using a physically damaged sensor – how could you do it? Imagine being able to extend the range

BROKEN SENSOR? NO PROBLEM!

Imagine being able to continue using a physically damaged sensor – how could you do it? Imagine being able to extend the range of a sensor 2 or 3 times, with no extra cost to your supplier? Imagine sensor processing tasks that are no longer limited to strict environmental conditions, sensor capabilities, or even to particular sensors…

With dynamically reconfigurable analogue circuits these scenarios are becoming a possibility. Of course the technical difficulty of designing analogue circuits hasn’t gone away, but we now have the software technology to allow us to discover analogue solutions that might take years for human engineers to discover. Each chip interfacing with a sensor can be reconfigured as many times as needed to ensure a linear response is met as conditions change. New circuits can be evolved in situ. Combinations of sensor inputs can be controlled and analysed by combinations of chips in circuit configurations that would have conventional analogue engineers in dismay!

Evolution allows this because it is a “test it and see” approach. No calculation is necessary. No engineering is carried out.

OUR INSPIRATION

It can be a shock to discover how often nature re-uses genes that make good solutions. Genes responsible for bone material, the formation of limbs, sight and a host of other morphological features are widely re-used in different contexts across species. The context in which a gene actives the development of morphological feature can affect its size (such as the decreasing size of vertebra in your spine), shape (such as appendages) or even function. But what is the mechanism for gene re-use? How is that certain genes are expressed in particular locations in the embryo and not others?

Developmental processes create repeated morphological structures due to the distribution of binding proteins throughout the embryo. The presence of proteins that bind onto sections of the DNA acts like a “switch” that determines which genes are expressed in which location. Thus DNA and its associated RNA transcription processes produce proteins that affect the production of other proteins later in the developmental process. In effect, DNA produces the rules that govern its own production of other rules and other proteins!

A similar mechanism is used to control the body’s responses to changes in the environment. The presence of a particular protein in a cell (such as an auxin inhibiter) may trigger a chain of other proteins being produced that lead to the cell undergoing mitosis.

BACKGROUND

Evolutionary computation has been hampered by an inability to tackle problems of significant complexity or scale. Although it has achieved notable successes, its application range has extended little beyond search-based optimisation tasks. To date, no one has evolved a car, a house or anything where the number of parts exposed to mutation and selection is large and interconnected. However, Thompson’s (1996) ground-breaking hardware experiments showed that evolution can be used to discover solutions in parts of the design space normally inaccessible to human engineers. Evolution is able to operate in design domains whose complexity and physical richness would defeat human engineers, who usually try to simplify or abstract away from physical reality in order to make their problems tractable.

RICH PHYSICAL SEARCH SPACES

The use of evolutionary algorithms to discover novel (even patentable) analogue circuits has a relatively long academic history. The evolvable hardware initiative from NASA also explored the innate complexity of physical search spaces. However, for the first time, we have the opportunity to explore a physical medium that is capable of continuous real-time reconfiguration. The new dynamically programmable Analogue Signal Processors (dpASPs) are destined to replace many ASICs (application specific integrated circuits), and offer potentially disruptive technology wherever digital logic has to interface with the real world.

These chips offer us the chance to return to the complex world of analogue signal processing with new eyes: eyes provided for us by evolutionary search algorithms.

Binding Analogue Signals To Digital Genomes: using feedback to guide evolutionary search

HOX GENE EXPRESSION DURING DROSOPHILA DEVELOPMENT

Stripes, spots and dashes: the presence of binding proteins distributed in varying degrees of concentration in individual cells

allows gene “switches” to be activated. Which genes are expressed (or inhibited) in turn leads to the context specific production of proteins at these locations, giving rise to repeated, modular

morphological structures such as body segments, wing disks, etc.

Credit: Dave Kosman, UCSD

Kester Clegg ([email protected])

Supervisor: Prof. Susan Stepney ([email protected])

http://www.cs.york.ac.uk/~kester

PATTERN FORMATION AND MODULAR RE-USE

Top: The presence of different binding proteins combine to produce spots on a drosophila embryo. Bottom: Different switches causing

the gene for Bone Material Protein 5 to be expressed in different locations in a mouse embryo. The resulting morphology is context

dependent.

Image taken from Carroll (2006:125)

THE TECHNOLOGY

The Jacyl AXR-16 board: Contains 4 Anadigm FPAA chips (shown on left) which can be daisy-chained together. The chips

use switched capacitors to reproduce analogue components. The reconfiguration of the chips can be controlled using the on-board

FPGA (larger chip on right) or by an externally attached PC. Each of the FPAA chips can accept up to 4 inputs and have up to two

outputs. Their reconfiguration can be independent of one another. An API for the reconfiguration software is provided by Andadigm.

AnadigmDesigner2

Anadigm provide software to be used as a GUI design tool for

testing and implementing analogue circuits on their chips. However, the software also has an API which

can be ‘wrapped’ in other code, thereby

automating the design process and

downloading of circuits.

EXTERNAL INPUT

OSCILLOSCOPE PROBES

Oscilloscope probes read the analogue signals output by the dpASP chips. The signal is

passed through an ADC and then matched

against the genome to see if the “signal” can bind to a part of the

genome. If it can, then the circuit is re-

configured according to the genome’s specification.

GENERATION OF THE PHENOTYPE: PRUNING DANGLING NODES

A visualisation of the graphs show that randomly generated wiring is haphazard, improbable, and often unusable! The top graph of 50 nodes and 8 dangling outputs needs to be pruned until a single

dangling output is left (if one chip is used) as shown in the bottom graph. Even very large graphs of a 100 nodes and more collapse

quickly when pruned, meaning that most of the genotype is not realised in the phenotype. However, this “junk” is still subject to mutation and may come into use in a later generation if the mutation is beneficial.

STATE-DRIVEN DYNAMIC RECONFIGURATION

State-driven dynamic reconfiguration works by continuously analysing the state of a chip. If a particular state is recognised, then the reconfiguration process is triggered. The Anadigm software only uploads the differences between the last circuit and the previous circuit, so reconfiguration is as fast as possible. A single CAM can have its parameters tweaked, or an entirely new circuit can be uploaded. In this picture, changing input from

the sensors triggers a reconfiguration of the circuit as needed.

METHOD AND PLATFORM

Our “evo-devo” platform consists of 4 FPAA chips daisy-chained together. Each of the chips can receive multiple inputs and each can be re-configured with a new analogue circuit independently. The reconfigurations can be controlled by the on-board FPGA chip or from a PC connected to the board. The genomes are first developed using the software shown on the left. Each genome is then “pruned” until the required numbers of dangling outputs are left. Dangling outputs can be fed into the next chip or to an external source. The phenotypes are then translated into circuit specifications. Circuits which are unable to fit on a chip (due to resource considerations) are discarded. However, circuits which have incompatible configurations (due to meaningless connections, different clock speeds, bizarre feedback ‘islands’ etc.) are only discarded if the Anadigm software prevents us from downloading the circuit. We are not too concerned by warnings from the software that these circuits cannot be analysed!

The mechanism of triggering the reconfiguration is unique, as it attempts to replicate the feedback process that DNA and binding proteins achieve to control further protein production. By hooking oscilloscope probes into the circuit, we hope to match the analogue signals with potential “binding sites” on the digitally represented genome. For this to be possible, the oscilloscope software passes analogue signals through an ADC and the signal values are fitted (via fast Fourier transform) against the bin values schema shown on the right.

Unlike most evolutionary computation, our genome is never fully realised at any one time. Instead, a large part of the genome contains “switches” that may or may not be activated on receiving a matching signal, leading to potential reconfigurations lying “dormant”, waiting for the right circuit conditions to appear. These switches are still exposed to the mutation of evolutionary processes and may be judged as part of the DNA “junk” if they are not used.

GENERATION OF GENOMES

Left panel: shows details of CAMs (Configurable Analogue Modules). Each CAM can be configured with different parameters, clock speeds and option settings to give different signal processing

capabilities. These configurable modules form around 168 building blocks that are represented in the genome shown on the

right. Right panel: using Cartesian Genetic Programming (CGP – developed by Julian Miller, now at York), a feed-forward directed

graph is generated with each node in a graph representing a CAM configuration. Feed-forward means that all inputs to nodes are guaranteed to be taken, but outputs can be left dangling and during the phenotype generation these need to be pruned.

NEUTRAL VARIATION

Evolution is remarkable as a search algorithm. Its ability to hill-climb is well known and has been used in many optimisation applications. However, it is as an algorithm of invention that it really stands out. When combined with rich physical environments, evolutionary algorithms frequently find solutions that are judged incomprehensible to human engineers. Adrian Thompson (1998) has suggested this is in part due to neutral variation in the genome allowing the algorithm to “jump” from peak to peak in the search landscape, as the neutral part of the genome is never realised or selected for. Evolution appears to be the only search algorithm capable of reaching solutions in this way.

THE WILDCARD BINDING SCHEME

We can only be inspired at how binding proteins attach to DNA and the complex chemistry involved. The mechanism of self-

regulation and context-specific response to environmental change is fundamental to how organisms are controlled.

In our model, the binding proteins become analogue signals with their matching binding sites on the genome taking a digital representation. In wildcard positions, any value for the converted signal will match, making the binding sites flexible in terms of what range of signals they will accept

and how they will act as switches. Analogue signal activity therefore has the potential to feedback into the genome,

triggering a reconfiguration when certain conditions are met.

HOW DO PROTEINS BIND TO DNA?

Proteins attach to stretches of DNA material by recognising signature sequences of base pairs. For example, a single switch for a gene may consist of several hundred base pairs (bp), lying perhaps several thousand bp upstream of the gene. Within the gene switch, there are usually 6-20 signature sequences (each 6-9 bp in length) that affect the expression of the gene concerned. Signatures are therefore very small, but this does not mean they are limited in their potential combinations (in fact, even a relatively short signature has a huge number of combinations). Signature sequences are sometimes exact for every position, sometimes they contain wildcards.

Wildcard positions can be filled by all four nucleic acids (Cytosine, Thymine, Adenosine, and Guanine) but are more often limited to pairs of alternatives (e.g T or A, C or G, etc.). For example, Tinman, a gene related to heart development in most species, from drosophila to humans, is highly specific (this example is taken from Carroll (2006)):

Tinman TCAAGTG

Pax-6 (eyeless) KKYMCGCWTSATKMNY

Dorsal GGGWWWCCM

whereas Pax-6 (the master gene controlling forms of sight in species), and Dorsal, use the wildcards represented by K (G or T), Y (C or T), M (C or A), W, etc. Thus Pax-6 has a signature with only 6 specific sites out of 16 possible bp combinations, meaning that it could bind at a variety of locations.

Our analogue signal binding scheme is shown below. It too has wildcards as bin values representing the transformed analogue signals. Like nature, we allow evolution to mutate these binding sites, allowing new “switches” to evolve that have different binding properties, allowing new genes to be switched on.