Recipe tuning by reinforcement learning in the SandS ecosystem

The Social and Smart (SandS) project ecosystem is compounded of household appliance users sharing recipes for the used of appliances, an intermediate control layer, and an intelligent social layer which aims to optimize the appliance recipes maximizing user satisfaction. We consider two aspects of the social intelligence, the innovation producing new recipes for unkown user tasks, and the adaptation to personalize the recipe to an individual user on the basis of his/her specific feedback. The second aspect is proposed to be dealt with by Reinforcement Learning approach, thus user feedback becomes the system reward. In this paper we discuss such an architecture based on the actor-critic approach, providing some experimental results on synthetic datasets that demonstrate the feasibility of the approach, previous to real life implementations.