A dynamic computational cognitive model can be used to explore a selected complex cognitive phenomenon by providing some features or patterns over time. More specifically, it can be used to simulate, analyse and explain the behaviour of such a cognitive phenomenon. It generates output data in the form of time series which can only be partially compared to empirical knowledge. This leads to a challenging problem to estimate values of the parameters of the model representing characteristics of a person. A parameter estimation approach for dynamic cognitive models is presented here by combining improved Particle Swarm Optimization (PSO) and Constraint Satisfaction (CS) methods. Having collected the key features of behaviour of a phenomenon, those are translated into a set of constraints with parameters that will be solved through an improved agent based PSO technique. Through this, within PSO each agent explores the complex search space while communicating the quality of a local parameter value vector relative to their current global best solution as a swarm (through cooperation and competition). This is performed in tournaments and results of each tournament are combined to address the premature convergence issue in PSO.