Machine Learning for Optimisation and Services
The MLS area at XRCE (Machine Learning for Optimisation and Services) addresses issues raised by the current, widespread trend in companies and organisations to adopt a service based business model. Service provision is a complex problem which has some generic aspects which call for generic solutions and methodologies. Beyond the need for an infrastructure in which to implement, deploy, monitor, maintain, upgrade and combine a set of services, which component architectures and Web services address, a crucial question is that of the optimisation of service provision, from both the provider's and the consumer's point of view, taking into account not only the knowledge of how a service works, but also the mass of information about its actual conditions of use. Mining such information is challenging, and using it to make the right decisions in the service provision is also a major issue. Our research focusses on these two topics, which correspond to two wide areas of research: machine learning and process control.
The main scientific challenge is to take into account a stochastic, possibly evolutive demand to make optimal decisions concerning the process, the effect of such decisions being itself only stochastically determined. In some cases, the optimisation may rely on a purely machine learning analysis to predict if a certain, usually unwanted state of a process is likely to happen given the current observations. Human controllers can be notified in such situations, and take appropriate corrective actions. In other cases, the problem may be cast as a pure stochastic control problem in which the most appropriate action is decided automatically (or semi-automatically with human supervision) taking into account the cumulated cost of the actions over time.
In both cases, the problem is translated into a mathematical optimisation problem in a possibly high dimension space and with a possibly complex objective function, usually an integral over a large number of stochastic outcomes ("max-sum" problems). Such problems are often untractable and unscalable, so that one essential direction of research concerns the development and analysis of approximate solutions (eg. mean field approximation and other variational methods). The optimisation problem becomes even more complex when the process may have an adversarial behaviour, ie. has the capacity to exploit flaws in the optimisation efforts of the process manager ("max-min-sum" problems). This typically happens when humans are involved in the process, e.g. when dynamic pricing is involved, and try to maximise their own profit by exploiting the knowledge of the optimisation policies. Another focus of research is therefore the design of mechanisms to prevent or minimise such behaviours while preserving an incentive for the agents to participate in the service game.