A forest dynamics simulation model, individual-based and spatially explicit, has been calibrated and quantitatively evaluated from data collected for more than ten years in the experimental plots of Paracou, in the rainforests of French Guiana. In the model, each tree is positioned on a virtual plot and follows rules life given by individual growth, mortality and recruitment sub-models according to a three years time step. The growth model is derived from a classical << potential >> << reducer >> form, adapted to the specifity of guianan rainforests. It predicts the diameter growth of a subject tree from its diameter at time t and two competition indices which describe its local environment : one of them takes into account the environmental pressure experienced by the tree at the beginning of its growth period and the other the dynamics of this pressure in the near past. The introduction of fifteen species groups in the model greatly improved its efficiency. Four mortality models predict standing mortality and primary, secondary or complex tree falling. A recruitment model allows new individuals more than 10 cm dbh to appear on 100 m2 square plots, depending on allowable space in the stand. The simulation model is programmed in SmallTalk-80, following the multi-agents system philosophy. A user-friendly interface allows : (i) to follow the evolution of a stand and of each of its tree components by means of maps and graphics ; (ii) to perform at any time different kinds of treatments in order to test silvicultural scenarios. Trials with the model show that virtual stands behave satisfactorily whatever the intensity of disturbances applied at the beginning of the simulation but also that its reactivity is too low as compared to a matrix model calibrated with the same data. We suggest different ways to improve it.