No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Evolution-In-Materio: Solving Machine Learning Classification Problems Using Materials
Abstract
Evolution-in-materio (EIM) is a method that uses artificial evolution to exploit the properties of physical matter to solve computational problems without requiring a detailed understanding of such properties. EIM has so far been applied to very few computational problems. We show that using a purpose-built hardware platform called Mecobo, it is possible to evolve voltages and signals applied to physical materials to solve machine learning classification problems. This is the first time that EIM has been applied to such problems. We evaluate the approach on two standard datasets: Lenses and Iris. Comparing our technique with a well-known software-based evolutionary method indicates that EIM performs reasonably well. We suggest that EIM offers a promising new direction for evolutionary computation.
... Here, we investigate how well the new reservoir/material framework compares in performance to the vanilla EiM technique, on a non-temporal task. The task is to perform classification on the Fisher Iris data set across a range of carbon nanotube composites, including a control case (a resistor array) and similar density materials used in the vanilla system [21], [22]. The final section presents two simple analysis techniques that could provide further understanding as to what reservoir/material properties are being exploited by evolution. ...
... The Iris data set 1 (also known as Fisher's Iris data set) is a well-known multivariate data type classification problem, and has been used to benchmark the vanilla evolution in materio technique [4], [21], [22]. The task is to classify three species of the Iris plant given the four attributes of petal and sepal length and width. ...
... Here the task is to classify binary classes, rather than a time-series output, so a threshold mechanism is used to translate the trained outputs into binary classes. To evaluate the accuracy of the binary classifier, and to conduct a fair comparison with the previous vanilla EiM results, the fitness calculation in [21] is applied: ...
The Reservoir Computing (RC) framework is said to have the potential to transfer onto any input-driven dynamical system, provided two properties are present: (i) a fading memory, and (ii) input separability. A typical reservoir consists of a fixed network of recurrently connected processing units; however recent hardware implementations have shown reservoirs are not ultimately bound by this architecture. Previously, we have demonstrated how the RC framework can be applied to randomly-formed carbon nanotube composites to solve computational tasks. Here, we apply the RC framework to an evolvable substrate and compare performance to an already established in materio training technique, referred to as evolution in materio. The results show that by adding the programmable reservoir layer, reservoir computing in materio can significantly outperform the original evolution in materio implementation. This suggests the RC framework offers improved performance, even across non-temporal tasks, when combined with the evolution in materio technique.
... Different organic and inorganic media have been used as materials, such as slime moulds [7], bacterial consortia [1], 978-1-5090-4601-0/17/$31.00 c 2017 IEEE cells (neurons) [22], liquid crystal (LC) panels [5] and nanoparticles [3]. Single-walled carbon-nanotube (SWCNT) based materials have shown the potential to solve computational problems [8], [12], [32], [15], [18], [23]. In [26] it is argued that non-biological materials make a better medium for unconventional computing exploration. ...
... The purpose of LCs is therefore to provide a medium within which SWCNTs are able to form variable and reconfigurable percolation paths resulting in complex electrical networks as argued in [11]. This adds an extra level of complexity and reconfigurability compared to the experiments performed on solid SWCNT composites, [8], [12], [18], [17], [16], [23]. ...
This paper presents a series of experiments demonstrating the capacity of single-walled carbon-nanotube (SWCNT)/liquid crystal (LC) mixtures to be trained by evolutionary algorithms to act as classifiers on linear and nonlinear binary datasets. The training process is formulated as an optimisation problem with hardware in the loop. The liquid SWCNT/LC samples used here are un-configured and with nonlinear current-voltage relationship, thus presenting a potential for being evolved. The nature of the problem means that derivative-free stochastic search algorithms are required. Results presented here are based on differential evolution (DE) and particle swarm opti-misation (PSO). Further investigations using DE, suggest that a SWCNT/LC material is capable of being reconfigured for different binary classification problems, corroborating previous research. In addition, it is able to retain a physical memory of each of the solutions to the problems it has been trained to solve.
... A number of recent investigations have demonstrated the capabilities of SWCNT/ polymer mixtures, in particular Poly-methyl-methacrylate (PMMA) and Poly-butylmethacrylate (PBMA), as a potentially rich, evolvable and ubiquitous substrate. These investigations include: solving classification and optimisation problems such as frequency classification [78]; classifying various data instances [23,80]; solving small numbers of cities instances of the travelling salesman problem (TSP) [22]; and applied to the (NP-hard) bin packing problem [79]. Early evidence highlights and supports the plausibility and potential of the methodology, but, in some respects it still lacks competitive results and still exhibits some of the issues raised at the end of §3.1. ...
... Simple classification problems are wide and varied in the field of machine learning, some of which can be seen in both the RC and EIM literature. For example, typical tasks for EIM are tone and frequency discriminators, and Iris & Lenses dataset classification [38,78,80]. Examples in reservoir computing include, signal classification (discriminating between two waveforms) [85], various n-bit parity problems [11,25,91,95] and other time-independent classification tasks [4]. ...
Research in substrate-based computing has shown that materials contain rich properties that can be exploited to solve computational problems. One such technique known as Evolution-in-materio uses evolutionary algorithms to manipulate material substrates for computation. However, in general, modelling the computational processes occurring in such systems is a difficult task and understanding what part of the embodied system is doing the computation is still fairly ill-defined. This chapter discusses the prospects of using Reservoir Computing as a model for in-materio computing, introducing new training techniques (taken from Reservoir Computing) that could overcome training difficulties found in the current Evolution-in-Materio technique.
... Liquid Crystals (LCs) from a display screen have been used as the material part of EiM for evolving robot controllers [7], a tone discrimination device [5] and logic gates [6]. In [10], [13], and [20] a dry mix of Single-Walled Carbon Nanotubes (SWCNT) with a polymer were used as the computational material and its electrical conductance was used as the manipulated property for solving the problem of calculating Boolean functions using a threshold interpretation scheme; the same material is used in [15] and [16] for solving optimisation problems. A mixture of SWCNT and LC in liquid form has also shown potential to solve classification problems. ...
... Liquid Crystals (LCs) from a display screen have been used as the material part of EiM for evolving robot controllers [5], a tone discrimination device [3] and logic gates [4]. In [9], [11], and [17] a dry mix of Single-Walled Carbon Nanotubes (SWCNT) with a polymer were used as the computational material and its electrical conductance was used as the manipulated property for solving the problem of calculating Boolean functions using a threshold interpretation scheme; the same material is used in [14] and [15] for solving optimisation problems. In [12], [22], [21] and [20] the material used is a solution of SWCNT with LC in liquid rather than in dry solid state. ...
... Друга иновативна идея е използването на еволюция в материали за адресиране на проблеми, типично адресирани от невронни мрежи. Група учени предлагат архитектура за решаване на проблеми на класификацията, за пръв път използвайки еволюция в материали за подобен род проблеми [2]. В друга публикация от 2013 е предложена невронна мрежа с емоции [4]. ...
In this paper key phrases in the field of neural networks are examined in terms of usage in recent scientific publications. For the purpose of the research several on-line academic databases are explored: Science Direct, Scopus, Google Scholar, Springer Link and others. The examined period is from year 2000 to present with focus on last decade. The terms are examined in two main directions: used paradigms and aspects of neural computing. Calculations determining the change rate of the usage of concepts and also calculations concerning the changes in their respective share in the papers on the topic are made. The data is used to draw some conclusions about the state-of-the-art of the field.
... Different organic and inorganic media have been used as materials, such as slime moulds [7], bacterial consortia [1], cells (neurons) [18], liquid crystals (LC) panels [6] and nano-particles [3]. Single walled carbon nanotubes (SWCNT) based materials have shown the potential to solve computational problems [8,11,22,14,15]. In [19] it is argued that inorganic materials make a better medium for unconventional computing exploration. ...
The potential of Evolution in Materio (EiM) for machine learning problems is explored here. This technique makes use of evolutionary algorithms (EAs) to influence the processing abilities of an un-configured physically rich medium, via exploitation of its physical properties. The EiM results reported are obtained using particle swarm optimisation (PSO) and differential evolution (DE) to exploit the complex voltage/current relationship of a mixture of single walled carbon nanotubes (SWCNTs) and liquid crystals (LCs). The computational problem considered is simple binary data classification. Results presented are consistent and reproducible. The evolutionary process based on EAs has the capacity to evolve the material to a state where data classification can be performed. Finally, it appears that through the use of smooth signal inputs, PSO produces classifiers out of the SWCNT/LC substrate which generalise better than those evolved with DE.
... Different materials that have been followed include biological material like slime moulds [10], [11], bacterial consortia [12] and biological cells (neurons) [13] as well as non-biological materials such as, liquid crystals [14], single-walled carbon nanotubes (SWCNT) [4], nano-particles [15]. SWCNT based materials have shown the potential to solve variety of computational problems [4] [16] [17] [18] [19]. ...
Evolution In Materio (EIM) is concerned with solv- ing computational problems by exploiting the physical prop- erties of materials. This paper presents the results of using a particle swarm optimisation (PSO) algorithm for evolving logic circuits in single-walled carbon nanotubes (SWCNT) based composites on a special custom made platform. The material used is a composite of SWCNT dispersed randomly in a polymer forming a complex conductive network. Follow- ing the EIM methodology the conductance of the material is manipulated for evolving threshold based logic circuits. The problem is formulated as a constrained, mixed integer optimisation problem. It is solved using PSO in conjunction with the shortest position value rule. The results showed that the conductive properties of SWCNT can be used to configure these materials to evolve multiple input/ output logic circuits.
... The nature of several computational problems requires more than a single output, e.g. TSP in [4] requires 9 to 11 outputs, classification in [20] requires two outputs, robot controller in [18] requires also two outputs. Let us consider a problem solved in-materio where two output values are required (for example the problem in [18], where a robot controller is evolved in-materio for controlling the speed of motor wheels to navigate a maze) and assume two different materials are tested: the first one with similar electrodes coverage and a second where one of the two output electrodes is barely covered by conducting material. ...
... In [5] and [4] liquid crystals (LC) panels were used for evolving logic gates, a tone discriminator and a robot controller. Single walled carbon nanotubes (SWCNT) based materials have shown the potential to solve variety of computational problems [7], [9], [23], [12], [13], [14] and [15]. ...
Evolution-in-Materio uses evolutionary algorithms (EA) to exploit the physical properties of unconfigured, physically rich materials, in effect transforming them into information processors. The potential of this technique for machine learning problems is explored here. Results are obtained from a mixture of single walled carbon nanotubes and liquid crystals (SWCNT/LC). The complex nature of the voltage/current relationship of this material presents a potential for adaptation. Here, it is used as a computational medium evolved by two derivative-free, population-based stochastic search algorithms, particle swarm optimisation (PSO) and differential evolution (DE). The computational problem considered is data classification. A custom made electronic motherboard for interacting with the material has been developed, which allows the application of control signals on the material body. Starting with a simple binary classification problem of separable data, the material is trained with an error minimisation objective for both algorithms. Subsequently, the solution, defined as the combination of the material itself and optimal inputs, is verified and results are reported. The evolution process based on EAs has the capacity to evolve the material to a state where data classification can be performed. PSO outperforms DE in terms of results’ reproducibility due to the smoother, as opposed to more noisy, inputs applied on the material.
This tutorial covers physical reservoir computing from a computer science perspective. It first defines what it means for a physical system to compute, rather than merely evolve under the laws of physics. It describes the underlying computational model, the Echo State Network (ESN), and also some variants designed to make physical implementation easier. It explains why the ESN model is particularly suitable for direct physical implementation. It then discusses the issues around choosing a suitable material substrate, and interfacing the inputs and outputs. It describes how to characterise a physical reservoir in terms of benchmark tasks, and task-independent measures. It covers optimising configuration parameters, exploring the space of potential configurations, and simulating the physical reservoir. It ends with a look at the future of physical reservoir computing as devices get more powerful, and are integrated into larger systems.
Modern computers allow a methodical search of possibly billions of experiments and the exploitation of interactions that are not known in advance. This enables a bottom-up process of design by assembling or configuring systems and testing the degree to which they fulfill the desired goal. We give two detailed examples of this process. One is referred to as Cartesian genetic programming and the other evolution-in-materio. In the former, evolutionary algorithms are used to exploit the interactions of software components representing mathematical, logical, or computational elements. In the latter, evolutionary algorithms are used to manipulate physical systems particularly at the electrical or electronic level. We compare and contrast both approaches and discuss possible new research directions by borrowing ideas from one and using them in the other.
Natural evolution has been manipulating the properties of proteins for billions of years. This ‘design process’ is completely different to conventional human design which assembles well-understood smaller parts in a highly principled way. In evolution-in-materio (EIM), researchers use evolutionary algorithms to define configurations and magnitudes of physical variables (e.g. voltages) which are applied to material systems so that they carry out useful computation. One of the advantages of this is that artificial evolution can exploit physical effects that are either too complex to understand or hitherto unknown. An EU funded project in Unconventional Computation called NASCENCE: Nanoscale Engineering of Novel Computation using Evolution, has the aim to model, understand and exploit the behaviour of evolved configurations of nanosystems (e.g. networks of nanoparticles, carbon nanotubes, liquid crystals) to solve computational problems. The project showed that it is possible to use materials to help find solutions to a number of well-known computational problems (e.g. TSP, Bin-packing, Logic gates, etc.).
Natural Evolution has been exploiting the physical properties of matter since life first appeared on earth. Evolution-in-materio (EIM) attempts to program matter so that computational problems can be solved. The beauty of this approach is that artificial evolution may be able to utilize unknown physical effects to solve computational problems. This methodology is currently being undertaken in a European research project called NASCENCE: Nanoscale Engineering for Novel Computation using Evolution. In this project, a variety of solutions to computational problems have been evolved using mixtures of carbon nanotubes and polymers at room temperature and also with gold nanoparticles at temperatures less than one Kelvin.
Evolution-in-materio (EIM) is a method that uses artificial evolution to exploit properties of materials to solve computational problems without requiring a detailed understanding of such properties. In this paper, we show that using a purpose-built hardware platform called Mecobo, it is possible to evolve voltages and signals applied to physical materials to solve computational problems. We demonstrate for the first time that this methodology can be applied to function optimization. We evaluate the approach on 23 function optimization benchmarks and in some cases results come very close to the global optimum or even surpass those provided by a well-known software-based evolutionary approach. This indicates that EIM has promise and further investigations would be fruitful.
Evolution-in-materio uses evolutionary algorithms to exploit properties of materials to solve computational problems without requiring a detailed understanding of such properties. We show that using a purpose-built hardware platform called Mecobo, it is possible to solve computational problems by evolving voltages and signals applied to an electrode array covered with a carbon nanotube–polymer composite. We demonstrate for the first time that this methodology can be applied to function optimization and also to the tone discriminator problem (TDP). For function optimization, we evaluate the approach on a suite of optimization benchmarks and obtain results that in some cases come very close to the global optimum or are comparable with those obtained using well-known software-based evolutionary approach. We also obtain good results in comparison with prior work on the tone discriminator problem. In the case of the TDP we also investigated the relative merits of different mixtures of materials and organizations of electrode array.
Evolution-in-materio (EIM) is the manipulation of a physical system by computer controlled evolution (CCE). It takes the position that to obtain useful functions from a physical system one needs to apply highly specific physical signals and place the system in a particular physical state. It argues that CCE is an effective methodology for doing this. One of the potential advantages of this is that artificial evolution can potentially exploit physical effects that are either too complex to understand or hitherto unknown. EIM is most commonly used as a methodology for implementing computation in physical systems. The method is a hybrid of analogue and classical computation in that it uses classical computers to program physical systems or analogue devices. Thus far EIM has only been attempted in a rather limited set of physical and chemical systems. This review paper examines past work related to EIM and discusses historical underpinnings behind such work. It describes latest developments, gives an analysis of the advantages and disadvantages of such work and the challenges that still remain.
Intrinsic evolution has been shown to be capable of exploit- ing the physical properties of materials to solve problems, however most researchers have chosen to limit themselves to using standard electronic components. However, these components are human designed and inten- tionally have predictable responses, so they may not be the most suitable physical devices to use when using a stochastic search technique. In- deed allowing computer controlled evolution (CCE) to manipulate novel physical media might allow much greater scope for the discovery of un- conventional solutions. Last year the authors demonstrated, for the first time, that CCE could manipulate liquid crystal to perform computa- tional tasks (i.e frequency discrimination, robot control). In this paper, we demonstrate that it is also possible to evolve logic gates in liquid crystal.
Although intrinsic evolution has been shown to be capable of exploiting the physical properties of materials to solve problems, most researchers have chosen to limit themselves to using standard electronic components. However, it has been previously argued that because such components are human designed and intentionally have predictable responses, they may not be the most suitable medium to use when trying to get a naturally inspired search technique to solve a problem. Indeed allowing computer controlled evolution (CCE) to manipulate novel physical media can allow much greater scope for the discovery of unconventional solutions. Last year the authors demonstrated, for the first time, that CCE could manipulate liquid crystal to perform signal processing tasks (i.e frequency discrimination). In this paper we show that CCE can use liquid crystal to solve the much harder problem of controlling a robot in real time to navigate in an environment to reach an obstructed destination point.
This paper presents a novel representation of Cartesian genetic programming (CGP) in which multiple networks are used in the
classification of high resolution X-rays of the breast, known as mammograms. CGP networks are used in a number of different
recombination strategies and results are presented for mammograms taken from the Lawrence Livermore National Laboratory database.
Thesis (Ph. D.)--University of Sussex, 1996.
Intrinsic evolution in evolvable hardware research has hitherto been limited to using standard electronic components as the media for problem solving. However, recently it has been argued that because such components are human designed and intentionally has predictable responses; they may not be the optimal medium to use when trying to get a naturally inspired search technique to solve a problem. Evolution has been demonstrated as capable of exploiting the physical properties of material to form solutions; however, by giving evolution only conventional components, we may be placing arbitrary constraints on our ability to solve certain problems. We have shown for the first time, that liquid crystal can be used as the physical substrate for evolution. We demonstrate that it is possible to evolve various functions, including a tone discriminator, in materio.
It is argued that natural evolution is, par excellence, an algorithm that exploits the physical properties of materials. Such an exploitation of the physical characteristics has already been demonstrated in intrinsic evolution of electronic circuits. This paper is an attempt to point the way toward the exciting possibility of using artificial evolution to directly exploit the properties of materials, possibly at a molecular level. It is suggested that this may be best accomplished in materials not normally associated with electronic functions. Electronic components have been prefected by human designers to construct circuits using the traditional top-down methodology. Workers in artificial intrinsic hardware evolution have with the best of motives, been abusing such components. It is a tribute to the amazing resourcefulness of a blind evolutionary process that it has been possible to evolve new circuits in this way. Artificial evolution may be much more effective when the configurable medium has a rich and complicated physics. This idea is discussed and particular examples that look extremely promising are given. Ultimately it may be possible to evolve entirely new technologies and new sorts of computational systems may be devised that confer many advantages over conventional electronic technology.
This paper presents a new form of Genetic Programming called Cartesian Genetic Programming in which a program is represented as an indexed graph. The graph is encoded in the form of a linear string of integers.
Evolution in Materio (EIM) exploits properties of physical systems for computation. “Designs” are evolved instead of a traditional top down design approach. Computation is a product of the state(s) of the material and input data. Evolution manipulates physical processes by stimulating materials assessed in situ. A hardware-software platform designed for EIM experimentation is presented. The platform, with features designed especially for EIM, is described together with demonstration experiments using carbon nanotubes in a thick film placed on micro-electrode arrays.
