Conventional information retrieval (IR) evaluation relies on static relevance judgements in test collections. These, however, are insufficient for the evaluation of interactive IR (IIR) systems. When users browse search results, their decisions on whether to keep a document may be infuenced by several factors including previously seen documents. This makes user-centred relevance judgements not only dynamic but also dependent on previous judgements. In this paper, we propose to use a probabilistic automaton (PA) to model the dynamics of users' relevance judgements. Based on the initial judgement data that can be collected in a proposed user study, the estimated PA can further simulate more dynamic relevance judgements, which are of potential usefulness for the evaluation of IIR systems.