Personalized Products Emerging from Tailored User Adapting Logic (Perpetual)

1 Jan 2012
31 Dec 2015

Combining an optimal product or system functionality with low resource consumption and minimal costs is a vision that fits well into recognized customer expectation trends. Conventional self learning and control strategies lack predictive capabilities to ensure a high functional service level in systems characterized by diverse usage patterns and user preferences, and in consequence do not provide effective solutions for achieving the envisaged resource efficiency. Traditional machine learning strategies typically also result in a significant risk of temporary discomfort as part of the learning phase, an obstacle that damages the acceptability towards end-users. Starting from experience built up in the domains of automatic knowledge profiling, reinforcement learning and advanced control, the partners in the proposed project consortium target to conceive methods and supporting algorithms to automatically determine (multi-)user characterizing demand profiles (so-called ‘usage profiles’) and derived statistical functional demand predictions that can greatly enhance the effectiveness and efficiency of self-learning control strategies. Based on time dependent monitoring of user demand signals and by using advanced datamining methods in combination with statistical time series analysis techniques, methods and algorithms will be developed that support the identification of multiple, distinct user demand profiles. An explicit scientific target is to develop a fully automated grouping (e.g. clustering) method that allows dynamic determination of an optimized number of distinct usage profiles in function of an improved control performance. These techniques will be explored for single and multi-user groups, with special attention for the influence of variable user group composition effects. The development of tailored decision models, compatible with the usage profile representations, is envisaged in support of functional demand prediction on a short (minutes, hours) to medium term (days) scale. The integration of the identified usage profiles into advanced, stochastic control algorithms will lead to a user-oriented system performance, Additionally, reinforcement learning will enable tracking and updating of the usage profiles over time and will moreover avoid discomfort during tuning of the product behavior, while contributing to the continuous minimization of resource consumption.