Relevance feedback (RF) is commonly used to improve the performance of CBIR system by allowing incorporation of user feedback iteratively. Recently, a method called image relevance reinforcement learning (IRRL) has been proposed for integrating several existing RF techniques as well as for exploiting RF sessions of multiple users. The precision obtained at the end of every iteration is used was a reward signal in the Q-learning based reinforcement learning (RL) approach. The objective of learning in IRRL is to estimate the optimal RF technique to be applied for a given query at a specific iteration. The main drawback of IRRL is its prohibitive learning time and storage requirement. We propose a way of addressing these difficulties by performing `pre-digestion' of concepts before applying IRRL. Experimental results on two databases of images demonstrated the viability of the proposed approach