Decompositional model-based learning and its analogy to diagnosis
A new generation of sensor rich, massively distributed autonomous system is being developed, such as smart buildings and recongurable factories. To achieve high performance these systems will need to accurately model themselves and their environment from sensor information. Accomplishing this on a g...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Text |
| Language: | English |
| Subjects: | |
| Online Access: | http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.498.50 http://www.aaai.org/Papers/AAAI/1998/AAAI98-027.pdf |
| Summary: | A new generation of sensor rich, massively distributed autonomous system is being developed, such as smart buildings and recongurable factories. To achieve high performance these systems will need to accurately model themselves and their environment from sensor information. Accomplishing this on a grand scale re-quires automating the art of large-scale modeling. To this end we have developed decompositional, model-based learning (DML). DML takes a parameterized model and sensed variables as input, decomposes it, and synthesizes a coordinated sequence of \simplest" estimation tasks. The method exploits a rich analogy between parameter estimation and consistency-based diagnosis. Moriarty, an implementation of DML, has been applied to thermal modeling of a smart building, demonstrating a signicant improvement in learning rate. A new generation of sensor rich, massively dis-tributed, autonomous systems is being developed, such as networked building energy systems, autonomous space probes, and biosphere-like life support systems, that have the potential for profound environmental and economic change (Williams & Nayak 1996). To achieve high performance, these immobile robots will need to develop sophisticated regulatory systems that accurately and robustly control their complex internal functions. To accomplish this immobots will exploit a vast nervous system of sensors to accurately esti-mate models of themselves and their environment on a grand scale. Handling these large scale model es-timation tasks requires high-level reasoning methods that coordinate a large set of traditional adaptive pro-cesses. Decompositional, model-based learning (DML) and its implementation Moriarty address this problem, providing a high-level reasoning method that generates and coordinates a set of nonlinear estimation codes, by exploiting a rich analogy to ATMS-based prime impli-cant generation(de Kleer 1986) and consistency-based diagnosis (e.g., (de Kleer & Williams 1987)). |
|---|