[Versión en Español]


Schema Theory

There have been a number of attempts to define a methodology for the analysis of complex dynamic biological systems. One of these attempts is schema theory which lays down the conceptual framework for knowledge representation inspired from biological and cognitive studies. Schema theory (Arbib 1992) contributes to Distributed Artificial Intelligence (DAI), and relies on cognitive science and brain theory. Its applications are in areas such as robotics, where cognitive, sensory, and motor processes can be represented at appropriate levels of detail. Schemas provide a level of representation at a "neuropsychological" level intermediate between gross neural task descriptions and detailed neural networks.

There have beed developed a range of schema theory based models, ranging from biological systems, such as models addressing lesion data on the toad's prey-acquisition and predator avoidance systems (Cobas and Arbib 1990; 1991; Arbib and Lee 1993; Corbacho and Arbib 1995), to models based on artificial neural networks, such as sensorimotor integration of robotic systems (Fagg et al. 1992).

A number of simulation systems have been developed to tackle different aspects of schema theory. In particular, ASL - Abstract Schema Language, is currently being developed to provide a modeling platform for neural based schema systems.

Adaptation and Learning

The area of learning and adaptation in dynamic systems has been addressed by many scientists in the traditional area of artificial intelligence, with systems such as SOAR, and in neural network systems, with biologically inspired learning such as hebbian , or artificial learning such as with back-propagation. Learning in traditional artificial intelligence systems is explicit, through pre-established high level mechanisms, while learning in neural modeling is implicit, generally specified as a mechanism at the level of individual synaptic weights. In traditional neural networks, the system has no way of knowing or reasoning about its state, as opposed to the more traditional explanation-based reasoning, thus constraining any optimizations or analysis which may be done to higher levels of cognition. We are yet to see evolved models where cognitive learning, (i.e., learning at a high level), is causally connected to learning at lower levels. Adaptation and Learning at multiple levels is one of the research goals in the laboratory.

Reflective Meta-Level Architectures

Reflective meta-level architectures add "introspective" capabilities to computational systems by providing dynamic adaptability to internal and external processing constraints [Smith 1984]. Reflection in object-oriented systems [Maes 1987] separates base-objects from their corresponding meta-object views, such as operational, resource, statistical, and migrational, where modification to any meta-object would cause a corresponding base-object modification, and similarly the other way around [Okamura et al. 1992]. Benefits of "exposing" system implementation in order to provide greater expressiveness at a more abstract level [Kiczales 1992] are seen in areas such as parallel systems, where greater efficiency in parallel compilers can be achieved by "opening" both the compiler and run-time system [Lamping et al. 1992]. Similarly, in the area of neural networks, a reflective architecture design can provide more adequate control of modular neural networks [Smieja and Muhlenbein, 1992], as well as improve the computational efficiency of the network [Boers and Kuiper, 1993].
In order to manage the inherent complexity in the development and simulation of computational models at multiple levels of granularity, we are currently developing a reflective meta-level architecture for NSL (Neural Simulation Language) and ASL (Abstract Schema Language), to provide the following mechanisms:

1. Provide processing (load balancing) and communication monitor and control capabilities in distributed NSL/ASL architecture.
2. Provide tracking capabilities for different ASL/NSL schema and neural network model and data versions.
3. Provide real time monitor and control capabilities when linking to real robots.