In this paper, we described how a machine-learning methodology could be applied to APM. Despite the simplifying assumptions we made about observability and temporal dynamics, the performance of our approach is better than naïve classification techniques and commercially available methods. Our success is partially due to the fact that we modeled the user context, which is a fundamental concept that arises in any autonomic domain. We are currently working on more complex models and on methods for automatic discovery of context and learning of temporal dynamics.

We believe that stochastic models such as MDPs, DBNs, and POMDPs are promising techniques for mitigating the complexity and handling uncertainty for autonomic systems. The POMDP model in particular facilitates reasoning about uncertainty and information; it allows the application designer to control the level of abstraction needed and to choose the appropriate tradeoff between model accuracy and feasibility. Yet, the POMDP model can be solved, at least sub-optimally, using relatively simple methods. Thus, the main challenge in applying the POMDP model in autonomics is not computational; rather it is a modeling challenge, i.e., creating models that are simple yet effective for the application at hand. We are currently working on using stochastic models, and POMDPs in particular, for APM.

Handling uncertainty and hidden information is at the core of many autonomic systems. Whether the designer of the autonomic system chooses a direct approach or whether a probabilistic model is preferred, the explicit consideration of the user is crucial for the success of many such systems. We believe that many autonomic systems of interest would benefit from addressing uncertainty and unobservable dynamics directly.