Presently, the uncertainties associated with controlling domestic heating system are managed using rule of thumb or heuristic rule-based controllers. The problems associated with this are: lack of bespokeness and optimality of the control to each unique building, difficulty in comparing technologies due to inconsistent control quality (lack of generality) and the expense of developing controllers for new technologies. In this work, the problem of heating system control is generalised with the intent of developing a generic-optimal controller — one that can control any set of heat sources in any building optimally, alleviating the aforementioned problems. A hybrid intelligent system design methodology is applied in order to develop the (model predictive) controller resulting in two sub-tasks. First, acquiring a model of each heating system — identification must be carried out on-line. Second, delivering optimal control using the model, given constraints. The first is tackled by applying Echo State Networks (ESN’s), whose benefits are that they have universal approximation ability, on-line learning is a recursive linear regression problem (for which the solution even in low precision environments is well understood), that on-line learning can be easily achieved using real feedbacks (network stability is relatively easy to attain) and that they can be easily scaled to systems of varying complexity. The second is tackled by using a global, derivative-free optimiser — meaning that the controller may tackle mixed integer problems and incorporate arbitrary output constraints expressed as penalty functions. A theoretical third problem arises due to the interaction of the learning and optimisation components of the controller. A methodology for tackling this is given. When applied to a simulated monovalent heating system in an unoccupied house (in the absence of user disturbances) consistent control can be achieved. The effective rejection of user disturbances is an outstanding problem and is briefly discussed.