For current projects see the website of my research group, Evo-DesignOpt. Read on for a rather dated (circa 2006) research perspective and overview.

The context of my specialization in evolutionary computation and humanoid robotics is an interest in complex systems. My goal is, with my students, to design radically innovative systems that address the need for computational systems to be more life-like, that is: adaptive, self maintaining and robust. I deal with software, hardware and their interaction with activity in the real world.

My Evolvable Hardware project that automatically designs analog circuits can be viewed from this perspective. Its analog circuit designs are organisms within an ecosystem where a dynamically reconfigurable analog array, test signals and test harness define an environment in which they must prove themselves. The organisms exist as a population that is subject to the process of evolution. Evolution derives power from the mapping between genotype and phenotype, inheritance, selection and genetic variation. It guides the system to robustly adapt circuits that fulfil the specifications of the design as a side effect of survival in the environment.

The project addresses the need for innovative approaches to supplement human expertise in analog design. Evolutionary algorithms have already been used for component value optimization. My group addresses the design of circuit structure. The testbench is capable of expressing candidate circuits which are specified in terms of components that humans use as design building blocks: differentiators, adders, integrators, etc. It combines human abstractions that exemplify design principles with nature-like processes that force circuits to adapt to design goals in order to survive. This will potentially yield a powerful paradigm for handling the rapid growth of complexity in the analog domain.

For the last year, I have been a member of the MIT team of the CEARCH (Cognitive Enabled ARCHitectures) project (DARPA, IPTO, ACIP). Our team is a proposing new architecture that, in order to exploit innovative power management, memory access strategies and silicon technology, will be more complex and flexible than ever. Central and requisite to the radical nature of this vision are a programming-execution model and application-independent framework that are capable of adaptation and introspection. Explicit, static, a priori commitments regarding the hardware-software substrate, micro-architecture, run-time environment or programmer-system interface, in general, will fail to exploit the design potential of this architecture. The system must be able to observe its behavior, formulate assessments of its performance and choose improved strategies for resource management. It must have a living framework.

One class of naturally occurring systems that are introspective and adaptive to dynamic resource demands are economic markets. They can be viewed as interacting groups of learning, boundedly rational agents whose dynamic demands for goods, in response to prices, influence supply. Agents have access to differing amount of historical data and pursue different strategies. Only those that can accumulate wealth survive. Markets are robust. They react and absorb exogenous and endogenous shocks. They exemplify distributed mediation of many interacting, heterogeneous interests in a non-stationary environment. We plan to design an agent-based computational market to study how it could provide introspection and adaptation in the context of the CEARCH run-time environment. Introspection in an application-independent framework raises the issue of how much historical data to examine. One specific, ideal duration can not be determined before the system is run. The ideal duration is changing because it ultimately depends on continuously changing world activity in the form of a computer system's usage by its users. A particular focus of our market will be mechanisms that foster the continuous emergence of timely, superior strategies among agents with differing memory lengths.

A third project I have started deals with determining how to exploit polychromatic solid state white light (LED) technology in order to take advantage of how much less power it expends versus incandescent or fluorescent lighting. The power requirements of white light in offices and large retail spaces in the USA alone cost billions of dollars. In anticipation of the technology advances that will lead to white LED adoption in the next 15 years, and in order to accelerate adoption, a PhD student, Maria Thompson, and I are working toward defining a new perspective of lighting and space. We intend to demonstrate how lighting architecture can be advantageously synthesized with physical spaces that are enhanced with awareness of their occupancy and contents. We recognize that people will always want light to be unobtrusive. We will not change this. However, we expect that they will question how they ever managed when lighting didn't invisibly modulate itself to provide them with better object rendition.

Our concept fundamentally relies on the quantitative measurement of human visual perception of color rendition. We are ready to start human factors testing to acquire such data. The concept also incorporates a vision of sensor networks to provide occupancy and object information. It also exploits how different mixtures of color constituents of white light have different energy profiles. We have plans underway for a software simulation of a simplified typical kitchen equipped with ceiling panels of white LEDs. The simulation will access sensor readings (reflectivity profiles) of objects in the kitchen and of occupants' activities. It will transition the room's lighting through imperceptible combinations of color constituents. It will act to save as much energy as possible when this is warranted by low occupancy or specific object color properties. It will provide the best lighting possible to enhance occupants' visual perception in the room. The basic design of the system will use a distributed, local-reference multi-agent model for the LEDs and sensors. A key question is how control algorithms can be designed to allow the system to meet its operational goals. We might possibly have to search and adapt control strategies via evolutionary algorithms. We may employ market mechanisms such as pricing and study their impact on the dynamics of power usage and lighting delivery. We will aim for the system to be robust in terms of self-maintenance as lights fail or degrade in performance. In short, we envision designing another living system capable of actuation and sensing of the physical world.

Finally, I continue to pursue robotics in a quest for living machines. Robots are embodied and situated systems that emphasize the nature of intelligence in terms of behavior rather than cognition. I am co-investigator on a NASA grant for a project on Autonomous Manipulation Capabilities. I am in the process of developing a mobile manipulation robotic platform that supports a special purpose behavior-based control language. This language supports a paradigm where the robot is a robust creature with incremental layers of 'younger' competencies which depend on 'older' lower layers. The paradigm assumes the robot is intrinsically coupled with the physical world via its morphology and this coupling allows its self-maintenance and adaptation to environmental unknowns.

In summary, my research enterprise can be described as the design of complex systems that integrate computational substrate and the physical world with living systems as their models.