Professor Ozalp Babaoglu: “self-* properties of complex systems“.
Current information systems have reached a level of complexity that makes it very complicated to manage and deploy distributed software using traditional techniques. Nowadays, most costs are related to maintain and fix existing systems rather than buying new equipment. To lower the total cost of ownerships of IT systems, humans should be removed from most operations that could be automated, therefore there is huge need for such system to posses self-{configuring, optimizing, healing, protecting, managing, etc…} properties (denoted self-*).
Prof. Babaoglu
In the blueprint for autonomic computing of IBM (quote), it is suggested that self-* would require an intelligent control loop. However, an alternative would be what he calls “grassroots” approach, which consist to interconnect many agents (rather than having a central intelligent control loop) where needed functionality would be an emergent property of the system. Emergence is unavoidable and is found everywhere: power grids, telephone switching nets, retail supply chains. The whole point is to control the interactions between agents such that the global behaviors which emerge are the ones that we want. Actions based on local information (swarm intelligence), with small number of components (I guess he means small as in relation to all components in the system)?
Prof. Babaoglu proposed the gossip method where interactions only happen with known peers, and all peers act identically. Peer sampling service of nodes random from the whole population. Overlay networks that satisfies desired topological properties: scalable, robust, decentralized, self-organized cell formation, some cells like or dislike each other.
He also discussed about the phenomenon of synchronization (eg. heart beats) and classical modeling using coupled oscillators, and how these can be modeled using the gossip framework – state (phi of the oscillator) that is communicated to a small set of neighbors. Finally, he proposed a model for formation creation (devices that self-arrange in a 2D ring formation) in a classic ubicomp scenario (does such a thing exist?), basically an ad-hoc network of mobile devices formed based on physical proximity and with multicast communication abilities.
My personal comments
This was a nice summary of the basics of swarm intelligence, with some insights on coupled oscillating systems that synchronize. Having worked for a while in both these fields, I would have been very happy to see more applicable description of these systems in the particular context of physical computing and actual ubicomp scenarios. In particular, I would have been glad to know more about (if any) attempts to formalize such complex systems (especially useful for verification and evaluation) and how to prove that a given (physical) system can behave deterministically in a certain way in a given amount of time (thus could be actually applied to concrete industrial settings having hard real-time constraints).
Also, I would be curious to know what are the global effects of using physical devices as opposed to simulated ones – can the emergent global behavior still be guaranteed with noisy sensors and communication channels, failing hardware, and so on? How would the wide range of possible failures be incorporated in more formal model for complex systems? These are important topics that are worth studying as I am sure that even if the system as a whole is robust to individual nodes “die”, I doubt it is also robust when the devices behave inappropriately.
Finally, as a question in the public mentioned, ubicomp scenarios are heterogeneous by “nature”, yet the gossip model presented assumes that all devices are identical and have the same properties/capabilities, which clearly limits the applicability of these models outside computer simulations. This is a very important question to my current research, how to efficiently and smoothly use a dynamic network of devices that can have very different capabilities (be it computational, sensing, communication, you name it)? I really look forward for new work in this direction.
