Islands of Near-Perfect Self-Prediction

In the course of e orts to more fully utilize the power of active networks to build a self-managing communications network, the nature of entanglement and the relationship between modeling and communication become of utmost importance. This paper provides a very brief introduction to Active Networks and the Active Virtual Network Management Prediction Project whose goal is a self-managing communications network. The focus of the paper is upon the e ects of near-in nite resources; that is, how will such a self-predictive system behave as processing and bandwidth become ever larger and more powerful. An attempt is made to identify new theories required to understand such highly selfpredictive systems. 1 ACTIVE NETWORKS AND SELFMANAGEMENT The problem this paper addresses is the complexity of managing large and rapidly growing communication networks with new more powerful technology. Network management consists of a wide variety of responsibilities including con guration management, performance management, fault management, accounting management, and security management. A network management system must be able to monitor, control, and report upon the status of all of these areas. Today, this is usually performed using a standards based management protocol such as the Common Management Information Protocol (CMIP) or the Simple Network Management Protocol (SNMP 1991). A goal of network management is to pro-actively detect problems in each of these areas. This means detecting potiential events such as performance problems and faults before they occur. This is a primary goal of Active Virtual Network Management Prediction (AVNMP) (Bush 1999). Active networks (Tennenhouse 1997) are a relatively recent concept in communication networks. Active networks are capable of executing general purpose code within packets as the packets are transmitted through intermediate network nodes. A framework for supporting the execution of general purpose code within packets as they travel through a network is an on-going research e ort. Thus active networks di er from today's communications networks because active networks o er a computational service in addition to a data transport service. In current communication networks non-executable data is passively forwarded through the traditional communication layers; intermediate devices such as bridges and routers only access the data link or network headers of packets. In active networks, intermediate devices can execute generic code within active packets as they travel through the network. The ability for communication networks to perform such computation o ers opportunities for great advantages in such areas as e ciency, rapid protocol development and deployment, and network exibility. However, active networks also add additional complexity, particularly in network management and security. The goal of Active Virtual Network Management Prediction is to use the advantages active networks provide in order to handle the additional complexity in network management. A detailed discussion of AVNMP is outside the scope of this paper, however, AVNMP's relevant characteristics are that it attempts to become a closed predictive system by injecting a simple model of the open ends of the network back into the network as shown in Figure 4. Another characteristic is that AVNMP attempts to utilize the inherent parallel nature of communications networks in an optimal manner, by using a Time Warp like mechanism to ensure causality among virtual messages within the system. The network uses information thus generated about its likely future state to improve its current performance. 2 NEAR-INFINITE RESOURCES Now, imagine stepping across a discontinuity into a world where computing power, bandwidth, and computational ubiquity are nearly in nite. Our vision focuses on e ects that near-perfect self-prediction would have upon such a world. First we would have near-perfect optimization of resources since local minima could be pushed far into the horizon. Second, currently wasted e ort could be avoided, since the outcome of any action could be determined with very precise limits. Critical missing elements are a theory and applications involving highly predictive systems and components. Study is needed to explore the exciting new world of near-perfect self-prediction and the relationship between highly predictive systems and communications in particular. Figure 1 shows an abstract view of computers embedded within almost all devices. Current engineering organizes computing devices in such a way as to optimize communications performance. In our hypothetical world of near-perfect predictive capabilities, direct communication is less important and, in many cases, no longer required, as discussed later. Instead, computational organization is based on forming systems or islands of near-perfect self-prediction. As shown in Figure 2, selfpredictive capability is used to enhance the performance of the system, which in turn improves the predictive capability, which again improves the performance of the system, ad in nitum, driving the error towards zero. Figure 2: This predictive capability is used to drive the error toward zero. Why do we assume near-perfect prediction rather than perfect prediction and why do we assume islands rather than perfect prediction everywhere? Clearly, perfect prediction everywhere would take us into a deterministic world where the nal outcome of all choices would be known to everyone and the optimal choice could be determined in all cases. In this project it is assumed that limits, however small, exist, such as lack of knowledge about quantum state or of the depths of space. In order to study near-perfect self-predictive islands, the characteristics of such islands need to be identi ed. It would appear that closed self-predictive islands would be the easiest to understand. The scope of closed self-predictive islands includes all driving forces acting upon the system. Imagine that one has full knowledge of the state of a room full of ping-pong balls and their elasticity. This information can be used to predict the position of the balls at any point in time. However, one is external to the room. The goal is for the balls to predict their own behavior as illustrated in the inner sphere of Figure 3. If elasticity represents the dynamics of communication endpoint entities A and B, and movement of the ping-pong balls represents communication, then any exchange of information between A and B is unnecessary since it can be perfectly predicted. Instead of transmitting messages between A and B, an initial transmission of the dynamics of A and B are transmitted to each other, perhaps as active packets within an active network environment. Thus a near-perfect selfpredictive island is turned inward upon itself as shown in Figure 4. In an active network environment, an executable model can be included within an active packet. When the active packet reaches the target intermediate device, the load model provides virtual input messages to the logical process and the payload of the virtual message is passed to the actual device, as described in the Active Virtual Network Management Protocol (Active Networks) project [DARPA ITO F30602-98-C-0230] (Bush 1999). A streptichron is an active packet facilitating prediction in our self-adjusting Time Warp System. GECRD is currently experimenting with streptichrons in an active network environment and this proposal takes that idea as close as possible to the extreme limit. Open self-predictive islands will contain inaccuracies in prediction because, by de nition, open selfpredictive islands include the e ects of unknown driving forces upon entities within the scope of the system. Figure 3 shows a force (F1) acting on the inner system. F1 is external to the inner system because it is not included within the system itself or in the virtual messages passed into the system. The system could become closed by either enlarging the scope to include the driving forces within the system, as shown in the gure, or by accepting a level of inaccuracy in the system. Thus we can imagine many initial points of near-perfect self-predictive islands, each attempting to improve prediction delity by expanding to incorporate more elements. These are the islands of near-perfect self-prediction. Figure 1: Computational organization is based on forming systems or islands of near-perfect self-prediction. Figure 3: Self-predictive islands can improve prediction delity by expanding to incorporate more elements. Figure 4: Direct communication between A and B is unnecessary as the dynamics of A can be transmitted to B, allowing B to interact with a near-perfect model of A. Recursion is a recurring theme in this work. For example, assume that the inner near-perfect self-predictive island in Figure 3 is a wireless mobile communications system and F1 is the weather. Now assume that ubiquitous computing can be used to include weather observation and prediction, for example, computers within planes, cars, space craft, etc. The heat from the circuitry of the wireless system, even though negligible, could have an impact on the weather. This is known as the butter y e ect in Chaos Theory. In recent years the study of chaotic nonlinear dynamical systems has lead to diverse applications where chaotic motions are described and controlled into some desirable motion. Chaotic systems are sensitive to initial condition. Researchers now realize that this sensitivity can also facilitate control of system motion. For example, in communications, chaotic lasers have been controlled, as have chaotic diode resonator circuits (Aronson, 1994) (DiBernardo, 1996). Hence, studying the e ects of external forces controlling a chaotic system has become a very important goal and should be a subject for research. A fascinating perspective on the topic of nearperfect self-predictive islands is found in Reference (Hofstadter, 1980), w