Ensuring thermal comfort is an important goal in th e operation of any office building. Often these buildings are therefore controlled according to the predicted mean vote (PMV) model which is also adopted as the ISO 7730 norm. This model predicts t he mean thermal preferences of an average group of people that would satisfy the thermal comfort ne eds of about 80% of the occupants. PMV based approaches have some drawbacks. Firstly, they inten d to capture the comfort needs of average occupants rather than those of the specific users s haring a room. Secondly, they are not applicable to all environmental conditions [Humphreys and Nicol, 2002]. Thirdly, they require a large amount of environmental data whose retrieval is very costly d ue to the many sensors needed. And, fourthly, getting the precise values for the required person- dependent data is often difficult. This paper prese nts two approaches that tackle all these issues. They b oth take previously recorded individual comfort votes of occupants into account and thus address th e first two weaknesses of PMV based approaches. The new average comfort vote is then predicted base d on the relevance of these existing votes to the new state. The two algorithms distinguish themselve s in the way they compute the degree of relevance of existing votes. One takes all the para meters of the PMV model into account in order to obtain a comfort vote prediction with high accuracy . The other one is only based on the temperature readings and thus, additionally, addresses the last two drawbacks of PMV based approaches. Both algorithms resort to a database DB in which ea ch entry