CAM-BRAIN : The evolutionary engineering of a billion neuron artificial
brain by 2001 which grows/evolves at electronic speeds inside
cellular automata machines.The states of cellular automata (CA) cells can be stored cheaply in RAM, so too the CA state transition rules. By using state-of-the-art CA machines (e.g. MIT's CAM8 machine, which can update 200 million CA cells per second), it becomes possible to grow/evolve neural networks based on cellular automata. Since giga-bytes of RAM are not too expensive, and the development of "superCAMs" (i.e. Cellular Automata Machines) which are thousands of times faster than CAM8 are achievable within a few years, it becomes realistic to develop artificial brains with a billion artificial neurons by 2001. This is the aim of the CAM-Brain Project at ATR Labs in Kyoto, Japan.
Three dimensional CA based neural circuits are grown and evolved to perform desired functions, even though how they perform their function is not understood. By evolving thousands of such circuits and their interconnections, a new field and probably a new industry called "brain building" may be born.
when
Artificial morphogenesis in optimization and compilationIn order to create systems made by many cells working in parallel, nature uses a developmental process. The development starts with a single cell which divides and divides again, generating a coherent parallel distributed system. We show how this simple idea of cell division can be exploited using computers, either for doing optimization or for compilation. In both case, the object being generated is a parallel distributed system.
when
Unconstrained evolution and hard consequencesArtificial evolution as a design methodology frees many of the conventional constraints normally imposed to make design by humans tractable. When evolving for hardware, we can relax such constraints as strict synchronization to a global clock, enforced decomposition into modules with simple interactions, and the use of high level abstractions such as Boolean logic, to name a few. However this freedom comes at some cost; there are a whole new set of issues relating to evolution that must be considered. Evolution is largely the dynamics of adaptation to a changed environment, which lends itself in artificial evolution to incremental increase in task complexity. Standard Genetic Algorithms are often not appropriate forthis, and need to be specially tailored. The main cost of an evolutionary approach is the large number of trials that are required. Attempted shortcuts through simulations raises further issues, and often robustness in the presence of noise or hardware faults is a crucial factor. To illustrate this a physical piece of hardware evolved in the real-world will be presented. A simple asynchronous digital circuit directly takes echo-pulses from a pair of left/right sonars, and drives the two motors of a real robot, so that it exhibits a wall-avoidance behavior. The complete sensorimotor control system (no pre- or post-processing) consists of just 32 bits of RAM and a few flip-flops, and is tolerant to single-stuck-at faults in the RAM. The rationale behind this experiment applies to many other kinds of system, including Field-Programmable Gate Arrays (FPGA's).
when
Development and evolution of hardware behaviorsA new system is proposed towards the computational framework of evolutionary hardware that adaptively changes its structure and behavior according to the environment. In the proposed system, hardware specifications, which produce hardware structures and behaviors, are automatically generated as Hardware Description Language (HDL) programs. Using a rewriting system, the system introduces a program development process, that imitates the natural development process from pollinated egg to adult and gives the HDL-program flexible evolvability. Also discussed is a method to evolve the language itself by modifying the corresponding rewriting system. This method is intended to serve as hierarchical mechanism of evolution and to contribute to the evolvability of large-scale hardware. Although this discussion is mainly involved in HDL-programs because our goal is hardware evolution, the techniques described are applicable to ordinary computers programs written in such conventional formats as "C" language.
when
Evolvable hardware with genetic learningThis paper describes Evolvable Hardware (EHW) and its applications to pattern recognition and fault-tolerant systems. EHW can change its own hardware structure to adapt best to the environment whenever environmental changes (including hardware malfunction) occur. EHW is implemented on PLD(Programmable Logic Device)-like device whose architecture can be altered by programming the architecture bits. Through genetic algorithms, EHW finds the best architecture bits which adapt to the environment, and changes its hardware structure accordingly.
Two applications are described: the exclusive-OR problem for pattern recognition and the V-shape ditch tracer with fault-tolerant circuit. First we show the exclusive-OR circuit can be learned by EHW successfully. This suggests that EHW may work as a hard-wired pattern recognizer with robust performance like neural net. The result is compared with neural net, classifier system, and adaptive logic network. The second application is the V-shape ditch tracer as part of a prototypical welding robot. EHW serves as backup of the control logic circuit for the tracing, although the EHW is not given any information about the circuit. Once a hardware error occurs, EHW takes over the malfunctioning circuit. The EHW architecture implemented on gate arrays is also described.
when
Embryonics: development of a new family of coarse-grained FPGA
endowed with the properties of self-repair and self-reproductionEmbryonics (embryological electronics) is a research project aimed at the realization of a new kind of electronic components which borrow three fundamental characteristics from living organisms: multicellular organization, cellular differentiation, and cellular division. These components will thus be endowed with properties heretofore restricted to living organisms: self-reproduction and self-repair. Within this framework, we will present a new family of coarse-grained Field-Programmable Gate Arrays. Each cell is a binary decision machine whose microprogram represents the genome, and each part of the microprogram is a gene whose execution depends on the physical position of the cell in the network, i.e., on its coordinates.
We will show a prototype of such a cell and use it to realize a cellular digital clock, capable of repairing and reproducing itself.
when
Embryonics: the birth of synthetic lifeThe field of Artificial Life is divided into three research axis. The first axis - Virtual Life - investigates simulation worlds (ants, worms, and so on...). The second axis - Alternative Life - addresses wetware developments of non-carbon-chain life. The third axis - Synthetic Life -synthesizes developmental and evolutionary concepts and applies them to engineering science.
Recent advances in the field of biology (evolutionary theory and developmental biology together with their engineering counterparts, genomic architectures and programmable devices) enable the birth of synthetic life. The Embryonics project is our humble contribution to this field of increasing interest. The project will be described as it developed since 1992. Both developmental and evolutionary VLSI will be described. Life-like properties as self-repair and self-configuration will be demonstrated and exemplified. Finally, open avenues and future developments will be presented and discussed.
when
Evolution and autonomous mobile roboticsAutonomous mobile robotics is a very promising but complex field. Autonomous vacuum cleaners, surveillance robots, automatic demining vehicles and many other large scale applications are included under this designation. But despite its name, this domain is very different from "classical robotics" that we are used to see in big factories, and very few applications are in use. The problem comes from the very large and robust autonomy that this kind of mobile robots need and the incredible complexity of the environment in which the robots act.
The robot that we are used to see in car factories has an autonomy restricted to a repetitive task executed in a very simple and limited environment.
On the contrary, autonomous mobile robots very often face an unknown world with a high degree of complexity (shapes, textures, colors...) and operate in a very large working area. The interaction between the robot, its control system and the environment in which the robot acts, play here a very important role. This has been clearly demonstrated by some examples proposed, for instance, by Franceschini, Dickmanns, Brooks.... However it is still difficult to design a control structure that fits very well with the hardware of the robot and to design a sensory-motor system that fits perfectly with the task and the environment in which the robot moves. In fact the exact characteristics of all these elements are often unclear or unknown.
The evolutionary approach can play an essential role at this level: the coevolution of the control structure and the hardware in the real environment under the control of a task supervisor provides a coherent solution. Waiting for an evolvable robot body, some experiments already show that the evolution of control algorithms in a conventional robot body generate a near-to-optimum exploitation of all sensory-motor possibilities.
when