2004”N“x XŠî‹àŒ¤‹†¬‰Ê•ñ
Œ¤‹†ƒe[ƒ}
l‚̈ʒu“Is“®“Á’¥‚ð”F’m‚·‚é’m“IƒG[ƒWƒFƒ“ƒg‚ÌŠJ” |
ŠT—v
In this research work, an intelligent computing environment is presented where a user's spatial behavior turns to be a trigger to retrieve and invoke ubiquitous objects within that user's living space.@For the computing environment to recognize a user's spatial behavior including his/her future movements, accumulations of GPS location information collected by a tiny program running on a user's cellular phone are organized into a structure which represents that user's spatial behavior. This structure provides computing environment with assumptions about the user's movement, and allows various ubiquitous computers to autonomously act for those inferred movements of the user. This structured knowledge representation is called "Behavior-based Personal Controller" or "BPC", since this structure could be a controller for invoking ubiquitous functions and retrieve ubiquitous information by the user's behavior. A learning algorithm to organize the Behavior-based Personal Controller and a prototype mechanism of a BPC to control networked ubiquitous objects are detailed in this paper. The accuracy of the learning algorithm is examined on two datasets of spatial behavior. |
¬‰Ê•¨
- T. Mizutori, K. Kohiyama, gRepresenting Human Spatial Behavior by Self-organizing Networksh, IFIP Intl. Conf. on Intelligent Information Processing, in October,2004.
- T. Mizutori, Yuta Nakayama, and K Kohiyama gBehavior-based Personal Controller for Autonomous Ubiquitous Computingh 2nd International Symposium on Ubiquitous Computing Systems (UCS 2004) ,Japan, November 8-9, 2004
Representing Human Spatial Behavior
For a comuting environment to understand the behavior of a user, a learning system of that user's geometrical movements was created. The user's location information is collected every 10 minutes through whole day by the tiny Java program running on the user's cellular phone (See picture below), and from the accumulations of the location logs, the learning system gradually organize the pattern of the user's geometrical movements or spatial behavior. | |
Location data collecting agent on the phone |
|
Learning System |
|
A learning network is constructed by a fixed number of nodes (called "behavior node"). A behavior node is also a three-element vector - (x, y, t), each element of which is initially given a pseudo random value ranging from 0 to 1. Behavior nodes are monotonically connected to represent, after learning, the sequence of behavior (i.e. transition of behavior) The learning procedure is based on the self-organizing neural network. For each log, every behavior node calculates the distance to that log. The distance between an input log vector I and a behavior node vector O, denoted as DIST(I,O), is calculated. After all the behavior nodes calculate their DISTs to the log, the behavior node with the minimum DIST value is selected as the "winner". The internal vectors of the winner behavior node and its neighborhood behavior nodes are updated. |
|
The analysis of learning performance was conducted on the simulator as below. |
|
The organized feature of the two subjects' week-day behavior.
Behavior-based Personal Controller
A figure below shows the 3D visualization of a BPC. The white connected nodes in the figure are the behavior nodes of the BPC. Tiny rectangle pictures linked to the behavior nodes are collected by the users' GPS-Camera embedded phones in order to create location based picture archives. Location-based picture archives are part of services called "moblog" where users' upload pictures with text notes through their camera embedded mobile phones. Here, these pictures express ubiquitous information related to the user's BPC. A sphere located at the center and linked to the behavior nodes around it in the BPC of the figure is a node which represents a networked table light (also shown in the left part of the figure), that is an example of ubiquitous computer. A networked micro web server "XPort" of Lantronix, Inc., an AVR microcontroller of ATMEL corp., and a solid state relay provide a power controlling logic of the ubiquitous computer (i.e. table light) from the BPC. In addition, IR transmitters are implemented on this networked controlling board to control remotely other appliances. One scenario the BPC and this controlling board draw is that the table light is switched on by sensing the user's approach to the laboratory, and information about today's lab work (e.g. task list) show up on the user's mobile computer. | |
Applications
Frontier services of Behavior-based Personal Controllers shoud be discussed in this section, although further development may need to realize all of these services.
|
Conclusion
In this research work, a knowledge representation of a user's spatial behavior is created in order to handle ubiquitous objects by the user's behavior. Since this knowledge of the user's behavior is seen as a "personal controller" for invoking ubiquitous functions and retrieve ubiquitous information, we call this knowledge representation of the user as "Behavior-based Personal Controller". This automatic mechanism to handle ubiquitous computations and information will be powerful solution in human and ubiquitous computer interaction. A Behavior-based Personal Controller is built from GPS location information collected by a tiny agent program running on a user's cellular phone. A process in this agent is collecting GPS location every ten minutes and another process constantly sends this GPS location information to a server where the BPC organizing agent learns the feature of the user's spatial behavior in the duration of one-day. A learning network consists of a fixed number of behavior nodes, each of which is a three-element vector (x, y, t). One location log (longitude, latitude, and time of log) is normalized and brought into the network. The network adapts itself by modifying the behavior nodes to reflect the input of each location log. After a number of location logs are input into the network, the learning network organizes the feature of the user's spatial behavior, and this representation is called "Behavior-based Personal Controller". Ubiquitous objects are related to near behavior nodes by their locations and time (or areas and time intervals). In the current implementation, a controlling mechanism is specifically defined from the user's position, his/her nearest behavior node, and corresponding ubiquitous objects invocation, so that general frameworks to describe service models which relate semantics of the BPC and ubiquitous objects are highly required. One problem in the current algorithm is to normalize the all location logs among the first set of logs. This means that the algorithm is not dynamic for newly coming logs after the learning is finished on the first set of logs. In addition, even if we modify the algorithm not to normalize data and to deal with absolute location and time values for managing newly coming logs, the regulation on the learn rate to correctly reflect changes of the user's behavior is not clear yet. Resolving this problem is the most important part of my on-going research. By resolving above problems - incorporating BPCs to wider and more general frameworks, and improving algorithm to be robust for more dynamic circumstances, I am motivated to realize the BPC as real services which enhance our activities in growing ubiquitous computing environments. |