Paper to presented at conference on Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications VIII, part of the International Symposium on Defense and Security, 12-16 April 2004 in Orlando, FL USA, Wireless Intelligent Monitoring and Analysis Systems Nina Berry, Donna Djordjevich, Teresa Ko Ben Coburn*, Stephen Elliott*, Brett Tsudama*, Melissa Whitcomb* Embedded Reasoning Institute Sandia National Laboratories, Livermore, CA. ABSTRACT The wireless intelligent monitoring and analysis systems is a proof-of-concept directed at discovering solution(s) for providing decentralized intelligent data analysis and control for distributed containers equipped with wireless sensing units. The objective was to embed smart behavior directly within each wireless sensor container, through the incorporation of agent technology into each sensor suite. This approach provides intelligent directed fusion of data based on a social model of teaming behavior. This system demonstrates intelligent sensor behavior that converts raw sensor data into group knowledge to better understand the integrity of the complete container environment. The emergent team behavior is achieved with lightweight software agents that analyze sensor data based on their current behavior mode. When the system starts-up or is reconfigured the agents self-organize into virtual random teams based on the leader/member/lonely paradigm. The team leader collects sensor data from their members and investigates all abnormal situations to determine the legitimacy of high sensor readings. The team leaders flag critical situation and report this knowledge back to the user via a collection of base stations. This research provides insight into the integration issues and concerns associated with integrating multi-disciplinary fields of software agents, artificial life and autonomous sensor behavior into a complete system. Keywords: Wireless sensors, intelligent monitoring, lightweight agents, team coordination, data fusion 1 INTRODUCTION This project demonstrates the concepts and implementation issues used to develop an intelligent monitoring and analysis proof-of-concept system to replace an outdated centralized master-slave system. The research team ideally wanted to investigate how modern distributed computing paradigms could be used to improve their current system functionality. They believed that advanced decision-making and autonomy were the keys to achieving an intelligence monitoring approach that was robust and flexible. To accommodate these interests the project research team investigated and implemented a lightweight agent solution that exhibited many of the characteristics of the traditional course grain multiagent paradigm. Each container in the facility would be embedded with a lightweight agent to control the localized monitoring process. The overall system would exhibit aspects of team coordination by linking together three minimal teaming behaviors of leadership, membership, and loneliness as part of the lightweight agent functionality. The lightweight agents would move through the different behaviors depending on the status of the environment and other agents. In this manner the behavior of the complete system would emerge from these behaviors to form a collection of teams that were easily dissolved with a simple command to reorganize. While the process of forming a team was predictable, the actual members and location of these teams remained virtual and unpredictable. The stability of the system was kept in check by the global control of when teams would form and by the local desires and goals of each agent to remain in the lonely state of behavior as little as possible. This decentralized decomposition of the global monitoring problem into smaller local pieces for each team leader gave the intelligent monitoring problem a unique localized flavor that was easily managed. The autonomy of each agent gave the containers the flexibility to respond to their given situation, while maintaining stability through their collective teaming behaviors. The resulting demonstration system illustrated a robust and flexible solution could provide intelligent monitoring that exceeded its centralized counterpart. Section 2 of the paper begins the discussion with a detailed overview of the customers’ current facility monitoring problem and how facilities are monitored in general. Section 3 * Students in the Embedded Reasoning Institute Internship Program. looks at what intelligence really means in the monitoring problem and discusses the analogy of agents and distributed sensor networks. Section 3.2 provides insight into proposed enhancement to the current facility components. The implementation details, functionality of the teams, and results are provided in Section 4 and Section 5. An overview of related background research is discussed in Section 6 and Section 7 looks at future directions. 2 FACILITY MONITORING PROBLEM Facility monitoring generally resides in a class of problems that are commonly characterized by three main security issues (1) perimeter or exterior protection, (2) internal protection, and (3) content tampering. This project explicitly dealt with security issues related to (2) and (3), thereby assuming a certain amount of confidence in the existing external perimeter protection for the facility. Based on this assumption the customers’ main concerns centered on issues related to internal environmental monitoring, stability of each container, and insider threats (i.e., unauthorized tamper of the container by company personal). To provide additional monitoring and security, the customer also equipped each container with a battery operated wireless sensor pack prior to storing it in the facility. The wireless sensor packs on each container provided the customer with a mechanism to monitor the immediate environment surroundings each container. The facility also contained a single base station (i.e., PC) unit to monitor the status of the individual containers in the facility. During a critical situation (i.e., significant change in sensor data) this base station unit reports this information to a human at the control center. The base station maintains logs of all sensor data received from each container and reports all information wirelessly to a control-monitoring center at a different location twice a day. The current facility operations represent a classical centralized (master-slave or server-client) monitoring approach, where a slave entity (i.e., container) reads its sensor data and wirelessly reports to the master (i.e., single base station) unit. The centralized data collection also illustrates traditional sensor network data fusion to a single point or synch node (i.e., base station) in the network. While this centralized solution, provided some adequate answers to the issues (e.g., environmental monitoring, stability of each container) related to internal protection it does not provide solutions for the concerns associated with the insider threat. To provide further insight into the current functionality of the different components of the facility and their monitoring concerns the remainder of this section will provide details of the components in this problem space. 2.1 Facility components details The facility is an enclosed building with adequate security to protect the external parameters of the building. The facility has one large entry for the insertion/removal of the container. The facility provides a controlled internal environment to maintain constant temperature, moisture, and light while protecting the contents of the one hundred containers currently stored in this building. Additionally, this facility has a regular doorway used by company personal to enter and exit the facility for general maintenance issues related to the environmental control equipment, base station, or replacement of sensor unit(s). When either door is opened it takes thirty minutes to restore the facility to the subscribed environmental settings. All instances of entry into the facility are logged into the system to provide explanations for the abnormal readings from the effected sensors. The facility contains two additional component types: a base station and a collection of containers. Each container is equipped with a wireless transmitter/receiver and a sensor suite containing three sensors (1) temperature, (2) moisture, and (3) light. Each wireless container unit is capable of transmitting/receiving within the boundaries of the facility walls. These combined sensor components receive their power from a non-rechargeable battery source contained in the complete sensor unit for each container. The containers are stationary after being placed in pre-specified locations within the facility. The base station is a stationary workstation computer located along the wall near the employee entrance. This computer is equipped with a wireless transmitting/receiving device permitting communication with the internal room. The base station also contains a larger wireless transmit/receiver antenna outside of the building to communicate with the external control room. 3 INTELLIGENT MONITORING While the customers had experienced adequate results with the centralized monitoring solution, they wanted to explore the potential incorporation of new advancements in distributed collaboration techniques to develop an ‘intelligent’ monitoring approach. This new approach would emphasize enhancing the internal protection and reducing the insider threat issues in their problem domain. While the artificial intelligence community initially coined the term ‘intelligence’ as a differentiating factor; they quickly found that the ability to define this concept was as illusive as the ability to concisely illustrate a convincing intelligent application. While the community long ago stopped arguing the exact definition of ‘intelligence’, the usage of the term has become more prevalent as distributed computing continues to spread across multi-disciplinary problem spaces. Hence the often overly used (often misused) concept of ‘intelligence’ is today used by many disciplines to signify the incorporation of advanced heuristics, statistical techniques, and/or decision-making capabilities into a given application. For this problem we will initially resort to the customers’ definition of intelligent monitoring as an application that exhibits distributed decision-making, while also incorporating aspects of individual autonomy. By using this initial definition, the research team approached this problem with a solution that emphasized the following three key factors decentralization, redundancy, and autonomy. Incorporating the concepts behind decentralization centered on problem decomposition to further simplify the global monitoring problem and distribution of the monitoring responsibility throughout the facility. This decentralization overlaps with the needs associated with redundancy, which provided robustness and flexibility to the solution. The key issue of autonomy was a predefined customer requirement, as well as, a necessity that would support individual flexibility, robustness, and multi-level decision-making. Based on the concepts associated with these three key factors the distributed paradigm that seemed ideal for this intelligent monitoring problem was the area of multi-agent systems. 3.1 Agent technology and sensor systems Our approach to agent technology deviates from the traditional applications (i.e., decision support or command and control systems), where agents are only used as high-level decision-making entities. We instead embraced the inherent analogies between agents technology (decentralized, autonomous, proactive, situated in an environment) and the distributed (decentralized) loosely connected independent (autonomy and proactive) entities, known as sensor nodes and wireless sensor units that were stated in the research of Boulis and Srivastava 1 and Hairong, et. al.2. Based on this analogy we proposed to embed smart behavior directly within each wireless sensor container, through the incorporation of a single software agent into the sensor suite for each container. The solution also increased the number of base station units within the facility and places a single agent on each base station unit. The agents in this solution exhibited the following behaviors and/or functionality of traditional software agent designs: Autonomous – software agents exhibit degrees of autonomy, with the ability of the software to make its own decisions based on a set of internal goals. Communicate – software agents use peer-to-peer communication between themselves and the different base station units. Proactive – software agents will exhibit functionality/behaviors that are not event driven to promote their autonomous behavior activities. In addition to these commonly known functionalities/behaviors the agents will be given the added ability to exhibit the socially oriented behavior associated with team formation and activities (see Section 4 for more details). This added behavior is commonly seen in the mobile robotic3 and agent organizational4,6 community, as a means of achieving coordination and collaborative between the agents. In this context, team behavior provides a mechanism to direct and control the decentralized global and local levels of problem decomposition that reside in the overall system. Team – software agents will cluster into social collectives, known as teams. Predefined global and localized rules & regulations govern the internal function of each collective. While this agent-based approach appears to mimic the traditional multi-agent paradigm, our system implementation is considered slightly unique in its usage of lightweight agents. Unlike traditional multi-agent architectures, the intelligent monitoring system requires that a team of agents functions in a resource (i.e., memory, power, etc.) limited environment on the actual sensor platforms. To accommodate these issues, we proposed and developed a lighter or smaller software framework (known as the embedded reasoning institute-framework (ERI-FW)6), which removed many of the traditional agent services and unnecessary functionality. However, the ERI-FW does support peer-to-peer communication and an XML messaging system using a reduced set of FIPA (Foundation for Intelligent Physical Agents – standards organization) agent performatives. The research associated with defining lightweight agents strives to maintain the collaborative and social aspects of traditional agent-technology in the most minimal manner. In this research, we are mostly interested in identifying what set of primitive behaviors best matches the capabilities needed for agents within the resource restrictive space of sensor systems. To accomplish this, the intelligent monitoring system uses a collection of simplistic artificial life like behaviors that collectively mimic the resulting local team behavior. A detailed discussion of these primitive team behaviors is given in Section 5. 3.2 Facility component enhancements To provide a more robust solution, the project team proposed some enhancements to the two main components within the facility. The first modification was to increase the number of base station units for this specific application to six, thereby addressing the concern of a single point of failure and accomplishing the first phase of redundancy within the system. Each base station would also require their own internal wireless transmit/receiver with the ability to monitor and log the entire internal facility from their locations. Each unit also has its own external antenna outside of the facility to transmit data back to the control facility. To resolve the coordination issues between the multiple base station units, the research team proposed to use a mirroring approach. This technique would treat the stations as a collective, where each unit is a mirror of the lead or original unit. The units monitor the health of the collective and are prepared to replace the lead base station if it fails to perform its duties. This replacement is done in a seamless manner that is not obvious to the containers or the control center; accept in the notification of the top base station failure in the logging files. The modifications to the containers included the addition of an individual motion detector to the existing sensor suite. This new sensor will be used to detect movement and tampering of the individual containers. In addition to this modification, a new restriction is imposed on the containers to prohibit the unauthorized movement of a container except by authorized personal with the capability of applying the necessary codes to deactivate the motion detector sensor prior to movement of the container. Violation of this new facility rule will cause the motion detector to signal an unauthorized movement of the container that will be logged and reported as a high priority. Figure 1 provides an illustration of the internal facility layout with the six base stations and one hundred containers. Figure 1: This diagram illustrates the facility containing one hundred wireless containers devices and six base station units. Each container is placed at a designated location within the facility and logged onto the facility map located on each base station. The large door at the base is for the insertion and removal of the containers. The small door at the right is for human investigation of the facility. 4 FACILITY MONITORING AND ANALYSIS SYSTEM The monitoring and analysis project resulted in a viable proof-of-concept system incorporating the lightweight multiagent teaming paradigm and the three key concepts presented in Section 3. Based on the proposed architecture this conceptual demonstration system illustrated the following collection of activities: (1) individual containers that selforganize into virtual random teams, (2) homogeneous lightweight agents that exhibited three primitive team formation behaviors (leader, member, lonely), (3) social teams locally monitoring and analyzing their environments, and (4) redundant motion detection that hinders the insider threat. While designing the intelligent monitoring solution, the customer was also concerned with another security issue associated with the eavesdropping of wireless communications within the facility. Once again the customer required reassurance that a third party could not simply determine a noticeable pattern in the team formations. To accommodate this requirement the containers self-organize into random virtual teams that are distributed throughout the facility. Hence a team leader could be located in the middle of the facility and its members be scattered at the perimeters of the facility. The reorganization of the teams is also done in a random manner, which proved to reduce the likelihood of the same team leader during the following reorganization to 25%. A detailed discussion of the teaming behavior that addressed concepts used to accomplish (2) and (3) is given in Section 4.1. The next major issue that concerned the customer was the need to prohibiting or hindering an insider threat attack. To understand the insider threat better, consider the following scenario where a trusted company employee enters the facility via the employee door using the appropriate entry code. This employee does not want to remove the container; instead they desire to simply remove the contents of the container, which would raise less suspicion. In this situation, the centralized approach could not have caught this intruder since the containers only responded to request from the single base station and there was no motion detector to prohibit tampering. While the intelligent monitoring approach cannot prohibit this employee from taking the container content, it will log the tampering of the container with every container in the facility. This redundant container level behavior only makes the process of keeping this tampering a secret harder, since each container must be rebooted manually in an attempt to erase their memory. The redundancy is also provides a backup for the base station units, which could have all been disabled as part of this break-in. Once the base-stations were brought back on-line each container is waiting to tell the control center of the unfortunate activities that have occurred. A quick cross-references of employee entry logs would permit the control center to determine the employee to question about this incident. 4.1 Teaming behavior The software agents embedded in each container are homogeneous discrete entities implemented with the combined functionality to become a leader or member. The humans in the control room govern the size, number and timing of the self-organization process that places the containers into teams. Each team will be dynamically disbanded and recreated whenever the facility is requested to reorganize or when a team leader is lost. Within the team, the responsibility of each software agent representing a single container is governed by a strict set of social rules used to hold the team in check. Figure 2 illustrates the three primitive behaviors for each individual agent, which is discussed in detail through the remainder of this section. These behaviors are part of the overall life cycle for each agent. An agent life cycle represents a circular chain of processes that software agents generally traverse during their live state. The life cycle for the lightweight agents used in the intelligent monitoring system are simple and consist of four simple stages: initialization (booting and/or rebooting), lonely, member, or leader. When each container is turned on, the embedded agent enters a life cycle that begins with a boot up process. This boot up behavior runs a set of normal power-on tests, then checks and recalibrates the sensors as needed. After the container finishes the boot up process, it enters a lonely behavior to begin the process of integrating itself into the rest of the network. First, the container needs to determine if there are leader agents that it should follow or whether the container should become a new leader. The container request a list of leaders from the lead base station and sequentially traverses the leader list until it finds a leader willing to accept its request to enter its team. If no existing team is found, the container becomes a new leader and begins the process of forming its own team. This approach will permit the formation of a single team consisting of one container. At this phase in the life cycle, the agent has successfully ended the lonely behavior. However, the agent will re-enter the lonely behavior if the team leader is lost, a request to reorganize is received from the control room, or the container is rebooted. Lonely Behavior Wait a random period of time (bounded) Ask if there open teams If yes, join available team as a member If no, become a leader Member Behavior Report data to leader Follow commands issued by leader Becomes lonely if it loses leader Leader Behavior Report information to base station Investigate abnormal data Follow/forward base station commandsReorganize request become lonely Figure 2: Three encapsulated primitive teaming behavior. Each begins life and re-enters the lonely during team re-organization. An agent resides in lonely state for limited time and quickly becomes a member or leader of a team. The member agent only initiates conversations with the leader agent, except for the motion detection command. The leader controls all team activities and investigates necessary abnormal situations. If the container becomes a member of a team, the embedded agent will enter the member behavior phase of the life cycle. During this behavior, the container reports sensor data to the team leader which performs preliminary analysis before forwarding this information to the base station. One potential scenario has this container detecting a high temperature reading and forwards this information on to its leader as an abnormal situation. The leader does not know if the high temperature reading is due to a malfunctioning sensor or an actual environmental condition. To assess this, the leader agent requests temperature sensor data from the other agents in physical proximity to the abnormal container. After collecting this sensor data, the leader agent determines that it is likely a sensor malfunction because the data from the other containers do not support the readings of the abnormal container. This information allows the leader agent to assign an appropriate low priority (yellow caution) status to the sensor data from the abnormal container. The leader agent also takes the proactive step of requesting the abnormal container to reboot so that its sensors will be tested and recalibrated. This completes one possible life cycle of a container agent, from hardware boot up to hardware reboot. If this container fails the reboot process a message is sent to the control center requesting a human to check the abnormal container. If the container becomes a “new” leader agent, it switches to the leader behavior within the agent life cycle. Even if the container does not receive any member agents for its team, the container will still be an active part of the network. As a team leader, the container reports its own sensor data to the base stations and waits for other agents to request permission to join the team. The container will accept new members on its team until it has reached the maximum allowed team size. As the lead container accepts new member agents into its team the lead container will start forwarding their sensor data to the base stations. Each sensor reading will be checked by the team lead to make sure that it falls within acceptable bounds. To investigate abnormal sensor reading, the lead container requests an updated list from the base stations of the containers that are in physical proximity to the member that reported the abnormal sensor reading. The lead container then requests current sensor readings from each of the containers that are in physical proximity to the container that initially reported the abnormal sensor reading. After the lead container gathers this information, it compares the responses of the other container to determine if they support the assumption that this sensor reading is from an environmental condition. If the abnormal reading appears to be accurate the lead container will report it to the base stations with a higher priority (red flag) status. The whole sensor reading analysis procedure presented in this section permits the agents to handle simple sensor malfunctions on their own without requiring human intervention. 5 PROOF-OF-CONCEPT SYSTEM DEVELOPMENT AND RESULTS The proof-of-concept system used to demonstrate the intelligent monitoring and analysis solution consisted of a collection of 10 Sharp Zaurus PDA representing the containers and a laptop PC representing the base station. The devices (PDAs and laptop(s)) communicated using 802.11b and the supporting proprietary ad hoc routing algorithm on the Sharp Zaurus units. The Sharp Zaurus PDAs contain a 206 MHz processor, 64MB of memory, running Linux QT operating system and supporting a version of PersonalJava. The container interface is illustrated in Figure 3 as an interactive graphical user interface that permits the user to manipulate the sensor controls in a real-time fashion on the PDA devices. When each PDA is initializing the interface behavior area shows ‘initializing’, which means the container is going through its boot up phase. Once it has successful booted the behavior switches to the lonely phase. At this time, the container is waiting to become part of the network. To become part of the network, the control room must accept its IP request which appears as a separate window on the control console. Once the container is part of the network, its behavior will eventually switch autonomously from lonely to either member or leader. Figure 3: The container interface with controls representing the temperature, light, moisture, and motion sensors. The user can manipulate the container by pressing the sensor control buttons, thereby driving the reading up or down. The interface also informs the user of the current behavior (either leader, member, or lonely) being exhibited by the container. The last item on the container interface is the motion-warning areas at the bottom of the screen. The appearance of this dot signifies the unauthorized movement of a container. The main interface window used to interact and observe the base station activities is shown in Figure 4. The top half of the window illustrates the IP addresses of the current containers in the facility. The right side of the top window contains buttons that emulate the ability of the control center to dynamically alter the global conditions of the room such as the ability to suddenly request a reorganization or the removal of a container. These changes directly propagate down to the facility floor (across the PDA units) and require the containers to react to these situations in a timely fashion. The lower half of the screen provides a reference map of the container that does not reflect their actual location on the facility floor. This map instead illustrates the relative connections between the different containers; the lines connecting the containers illustrate their actual team members. The leader of the team is shown as the entities with the triangle. The double circle container designates a container that has signaled an unauthorized motion. The three circles (red) illustrated the confirmation of a high priority warning of abnormal sensor readings is occurring. These three circles will remain until the sensor returns to normal readings. The square (yellow) is used to illustrate the unconfirmed sensor warning, which signals the control room with a yellow flag warning. Figure 4##: This is the main base station interface window; it illustrates a map of the containers in the facility at the bottom. The top of the interface shows the IP addresses for each container in the large left window. To the right are four buttons that give the user the ability to (1) reorganize the containers, (2) show team/group information, (3) remove a container, and (4) show sensor data. The interface shown in Figure 4 serves to further illustrate an outcome of the multi-agent approach used in this problem. The ability of the teams to provide not just sensor data but knowledge, obtained by pre-processed data prior to being sent to the control center, illustrates a significant improvement over the centralized system. With this information, the humans can make better-informed decisions with prioritized knowledge provided upfront. 5.1 Embedded Reasoning Institute Framework6 The Embedded Reasoning Institute Framework (ERI-FW) is a lightweight framework designed to support structured communication and distributed reasoning among devices within a network of sensor systems. The ERI-FW is an enabling technology that supports application development for distributed sensor systems. The methodology behind the framework is based on a hierarchical reasoning structure which supports three levels of data and control refinement among the devices within the framework. The assignments at each level are partially based on device functionality including size, weight, space, location, mobility, connectivity, and communication. The domain requirements and constraints provide further insight into the final device assignment of the framework for a specific application. Figure 5 illustrates the full three layers of the ERI-FW containing first level Base Stations — (a) laptop and/or (b) PocketPC (PDA); second level Microprocessors — PocketPC (c) with sensor packs and/or microprocessor boards without sensor packs; third level Sensors — (e) tethered sensor packs, (f) MICA units with sensor 6. Data flow within the framework is traditionally vertical from the sensor devices (most memory constrained) upward where additional intelligence and analysis is performed. However, horizontal data flow can equally occur across the first and second levels. The decentralized nature of distributed sensor networks requires a unique level of system control among the devices. While the ERI-FW does not require the developer to use software agents, the underlying peer-to-peer communication and messaging protocol can support this technology. This permits the framework to enable collaboration and coordination control among the devices through the incorporation of lightweight software agents 6. ## Due to the potential lack of color during the printing process, we have manually altered the map window to provide the reader with the ability to better understand the different entities on the map. Figure 56: The flow of data and control through the ERI-FW is illustrated as both vertical and horizontal concepts that flow throughout the entire system. This framework provides a peer-to-peer communication link between sensor suite with onboard processing and microprocessors PDAs along with a variety of different base stations. 5.2 Lessons learned and future research directions The decentralized approach used in this project requires direct communication between the distributed entities in the system. The known power limitation and the fact that communication is costly in any wireless application, would lead many to question the impact of this distributed system. However we believe, the lightweight agents and the agent teams assist in reducing the need for continuous direct communication between the decentralized containers. Incorporated into these situations is the fact that the containers are only required to share information during timed intervals or when an abnormal situation suddenly arises. Additional enhancements to the current system would further limit communication between the team leaders and member to concur with any collection of the following situations: Members report during original team formation process, Members report only when abnormal situation occur, Members report only when probed by a team leader, Team leader investigate abnormal situations by querying other team leaders, to get updated membership environmental information, Team leader checks with a subset of members to provide a status of overall team health. While we do not know the results of the above modification(s), we do know that the current system generates fewer messages than the customers’ original master-slave system (due to master driving the control process of the entire system). We could speculate that theses modifications might provide an additional drop in message by at least 25%. After developing the intelligent monitoring and analysis system the team reviewed the results with the customers to determine other potentially interesting issues and research directions to consider in future versions of this system. While the lightweight agent design successfully addressed the problem space, the customers wanted the ability to get even more functionality from the leader nodes. The flip side of this two-folded problem is the constraints on the memory size and power verses the homogeneous nature of the lightweight agent design. The more complex the functionality of the lightweight agents the more space they require to accomplish their goals. One potential solution is to have a few superleader nodes scattered through the room to provide the additional decision-making, as needed, for these nodes. These super-leaders could be additional units that reside on some containers in addition to the existing sensor packs. Another solution is to alter the current sensor suite hardware to include a micro-driver that would hold this super-leader code; the unit could access the micro-driver as needed. Other future research directions include experimentation associated with different team formation algorithms to compare the current first come first serve algorithm that is currently executed by the lightweight agents. We would also like to experiment with an algorithm to automatically decide when to reorganize the facility containers. 6 BACKGROUND RESEARCH The multi-disciplinary approach of the facility monitoring and analysis project has borrowed concepts and technologies from research in distributed systems, data fusion, decentralized control, team coordination, sensor networks, agent-based systems, and artificial-life (a-life) like primitive behavior. While the combination of agent technology and distributed sensor fusion is not a new concept the traditional approaches have dealt primarily with decision support systems 8,9, command and control10 for military sensor systems, and team oriented mobile robotic3,7 applications. However, recent research done in Boulis and Srivastava 1, Hairong et. al.2, and Hairong11 illustrates the latest combining of agent-technology and distributed sensor fusion in a sensor network known as active sensing. The active sensing is closely related to the concepts used in our intelligent monitoring project with the idea of using single agents on the different sensor nodes. The differences between the approaches reside in the types of agents and the overall functionality of the resulting systems. The active sensor approach uses mobile entities that move from node to node processing a given sensor network query. In contrast, our lightweight agents are stationary and more closely mimic the functionality of the traditional multi-agent systems. Our agents are also capable of coordinating their activities with localized team behavior instead of the propagated global problem approach of the active sensor systems. While the incorporation of agent-technology into sensor networks is not unique, the different perspectives and potential methods for achieving this task is a research window that still remains wide open. As distributed sensor networks continue to evolve, so will the different methods for achieving coordination and data fusion while enhancing flexibility and improved performance. The recent finding in Lesser 12 partially supports some aspects of our decision and directions in the intelligent monitoring task. Lesser states that ‘one approach from a multi-agent system perspective is to partition the environment into organizations, letting some agents become individual managers that represent aspects of the problem’12. While his approach is legitimately not as fine grained as our individual agents on each container, the results of applying organizational behavior will provide similar results. 7 CONCLUSION The benefit of using a multi-agent paradigm in conjunction with the three key factors of decentralization, redundancy, and autonomy has provided a solution to the intelligent monitoring and analysis problem that stretches across many different levels. By applying these concepts to this problem, the research team was able to see potential improvements for the human controllers in: investigation capabilities due to pre-filtering of information (classification of abnormal sensor reading as high or low priorities), productivity (controller no longer have to enter the facility for every little problem), and recovery from some failure modes without human assistance. The application of these concepts have also provided immediate understanding of how greater autonomy and decentralization of the problem space into smaller localized problems, and randomized self-organization and reorganization of the facility are all collectively connected to successful monitoring of the facility. By modeling system-wide goals as emergent properties of a set of discrete entities we were able to develop a system that is much more robust and scalable than one that achieves its goals by directly controlling all of the individual parts. This means that each lightweight agent does not have to talk to all the other lightweight agents before it can decide what to do nor does the system need a system-manager agent to give directions to all the other agents. (The “base station” and “control center” are “reporting pathways” even though they may occasionally initiate requests that have system-wide effects.) Because the lightweight agents act independently, the control traffic needed by the system is brought to a minimum. This minimization of control traffic reduces the congestion on the network and the energy requirements of the hardware modules that will “embody” these agents. Lightweight agents that are modeled as a discrete set of goaloriented behaviors and the transitions between them also make our system very flexible. Agents can be modified or extended simply by changing the properties of their behaviors or adding new behaviors. The behaviors themselves are dynamically loaded code, so they can be easily changed without requiring changes to the rest of the system. Because the overall system behavior emerges from the collective actions of individual agents, the system behavior can also be easily tuned by modifying the properties of agent behaviors. In short, this multi-agent system provides a highly configurable solution that can easily be extended to support new capabilities or problem domains. 8 ACKNOWLEDGEMENTS Sandia National Laboratories is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the United States Department of Energy’s National Nuclear Security Administration under Contract DEAC04-94-AL85000. The Embedded Reasoning Institute (ERI) is a research initiative at Sandia National Laboratories, CA., which contains a research team and student research internship program for graduated and upper division undergraduates. The ERI specializes in technology applications in wireless smart sensor devices. 9 REFERENCES 1. A. Boulis and M. Srivastava, “Enabling Mobile and Distributed Computing in Sensor Networks”, http://www.ee.ucla.edu/~boulis/phd/special_issue_ToC.pdf. 2. Q. Hairong, S. Iyengar, K Chakrabarty, “Mulit-Resolution Data Integration Using Mobile Agents in Distributed Sensor Networks”, Presentation at DARPA SenseIT project close out, Nov 5-8, 2002 http://www.ee.duke.edu/~vishnus/DARPA/oifusion.pdf. 3. Ashley W. Stroupe, Martin C. Martin, Tucker Balch, “Distributed Sensor Fusion for Object Position Estimation by Multi-Robot Systems”, Int. Conf. on Robotics and Automation (ICRA'01), 2001. 4. Kathleen Carley and Les Gasser, “Computational Organization Theory”, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence, ed. Gerhard Weiss, The MIT PRESS, Cambridge Massachusetts, 1999. 5. CMU Intelligent software Agent Lab, Agent Storm, http://www-2.cs.cmu.edu/~softagents/agent_storm.html. 6. Nina Berry, “Distributed Embedded Reasoning: Enabling Intelligence for Ubiquitous Devices”, Sandia National Laboratories Communicator Research Highlight, March 2003. 7. Matt Rosencrantz, Geoffrey Gordon, and Sebastian Thrun, “Decentralized Sensor Fusion with Distributed Particle Filters”, Proceedings of the Conference on Uncertainty in AI (UAI), Acapulco, Mexico, 2003. 8. Vladimir Gorodetski, Oleg Karsayev, and Vladimir Samoilov, “Multi-agent Data Fusion Systems: Design and Implementation Issues”, Proceedings of the 10th International Conference on Telecommunication Systems Modeling and Analysis, vol.2, pp.762-774, Monterey, CA, October 2002. . 9. Brahim Chaib-draa, Peter Kropf, Se´bastien Paquet, “A Teamwork Test-bed for a Decision Support System”, ", Proceedings of EUROSIM 2001: Shaping Future with Simulation, Delft, The Netherlands, 2001. 10. Plamen V. Petrov, Qiuming Zhu , Jeffrey D. Hicks , and Alexander D. Stoyen, “A Hierarchical Collective Agents Network for Real-time Sensor Fusion and Decision Support”, AAAI/KDD/UAI -2002 Joint Workshop on Real-Time Decision Support and Diagnosis Systems, Edmonton, Alberta Canada, 2002. 11. H. Qi, X. Wang, S. S. Iyengar, K. Chakrabarty, "Multisensor data fusion in distributed sensor networks using mobile agents", Information Fusion, TuC2-11-16, Canada, August, 2001. 12. Victor Lesser, “Experiences Building a Real Distributed Sensor Network”, Lecture notes CMPSCI791L: Sensor Networks, http://www-net.cs.umass.edu/cs791_sensornets/.