In a distributed virtual environment (DVE), the bottleneck of performance is the communication among hosts which keeps the consistency of virtual entities. The neural-network models of state vectors for unpredictable entities in a virtual scene are build in this paper, and the functional-link net to real-time prediction of their behavior is applied. Firstly, the principles and characters of functional-link net are introduced in this paper; secondly, after description of the traditional dead reckoning (DR) algorithm, an adaptive version of the algorithm based on functional-link net is presented; the network software architecture based on the algorithm is also given; finally, an example of the algorithm is given with experimental data, which shows the good performance of it.