ERCIM News No.47, October 2001 [contents]
Global Ambient Intelligence: The Hostess Concept
by András Lörincz
The Neural Information Processing Group of the Eötvös Loránd University, Budapest, launched a project 24 months ago to develop a general methodology for distributed collaborating goal-oriented experts (including humans) without using any assumption on synchronization. The expert community will have adaptive means for distributed computing, editing, and decision-making. Medical applications and home monitoring are in focus.
The research has originated by noticing the need for adaptive asynchronous information collection and control methods for distributed decision-making. Several experiences (see, eg http://www.eds.com/case_studies/case_arkansas.shtml) point towards this direction. Our model example is the complex case of Parkinson disease patients (PDPs) using Deep Brain Stimulators (DBS). Simpler systems can be derived as analogies of this case.
For PDPs, histories of individual patients show that sporadic experiences at different hospitals need to be collected to improve evaluation. Remote setting of the DBS, augmented reality with information collection and data management together, could form global ambient intelligence (GAI) for full support of the patients. Data management should meet constraints derived from Good Clinical Practice and could be seen as an extension of drug administration procedures. GAI can serve the patient, can help the doctor and can estimate risk. Adaptive tools for such (safety critical) applications have become available by our latest achievements in reinforcement learning (RL). Optimization of decisions on continuing data collection or executing an action can be termed as optimization of perception-action loops. Such challenging optimizations concern partially observed environments, being in the focus of RL research. We are not aware of any similar global approach on the field.
The project aims at defining the hostess, who is a goal-oriented agent with possibly adaptive and distributed subsystems. The hostess serves someone, eg a patient, a visitor, the owner of a home, or groups of those. For the interactions with people served, the hostess makes use of signals that are important for efficient human communication, including facial expressions, prosody, body talk, and behavioral patterns. In case of Parkinson disease patients global ambient intelligence needs to exploit:
A particular aspect of the project is that decision is achieved by combining the assessments of experts who may be located at different sites and may have different roles. The access to (medical) data may improve the decisions at the expense of real time, communication time, and computational costs. Risk-sensitive decision-making needs to consider pipeline operations, both in communication and in computation tasks. The project is just-in-time-research. We develop components under the assumption that Quality of Service will be available for internet networks in about two years.
State-of-the-art techniques have been adapted, developed, and are ready to use on the following fields: connectionist parallel algorithms, reinforcement learning for partially observable environments, parallel techniques for satisfiability problems, adaptive and robust control methods, tracking and probabilistic inference methods, internet technologies. The components are developed in a modular fashion. Some of the components are in use or are ready to use. Examples include text clustering and classification tools for search on the internet and other databases, tree validation, clustering and classification tools for XML documents, time series recognition and time series prediction tools, internet crawler and hostess technology in JAVA.
At each stage, the technologies we have adopted or developed undergo functional testing. For example, our internet crawler has been under testing. Medical doctors will help to improve graphical user interfaces. Two other components close to testing are:
Results on these components will be available in six months. CFR will be tested by its distributed editing capabilities for expert teams. We see no bottleneck for global ambient intelligence. Today, all essential technology components are within reach, including:
Sub-components of the project have potentials for other immediate applications.
Individual sub-projects have been and are supported by:
The figure is about our internet crawler that combines support vector machine and RL. Three graphs show the search patterns of the traditional breadth first crawler, the state-of-the-art context focussed (CF) crawler (NEC Institute), and our crawler. It can be seen that our crawler travels further and collects more information (given by the red circles). The efficiency of our crawler is almost an order of magnitude better than that of the state-of-the-art NEC crawler.