Our SMDM at time t consists of variables as follows:
Besides the above mentioned variables, the SMDM uses functions and parameters to describe the behaviors of the observed object and the system evolution.
The SMDM is essentially a dynamic Bayesian network, and the DBN structure can depict all casualties in the model.
Figure 5 depicts the DBN structure of the SMDM in two adjacent time slices.
However, when the state space is high-dimensional as that in the SMDM, the estimation variance will be very high.
Since the DBN structures of the AHMM and SMDM are deferent, the exact inferring process of the SMDM also differs from that of the AHMM.
To evaluate the performances of the SMDM and RBPF for recognizing the destination of the maneuvering agent in RTS Games, we design a typical game scenario.
2, we generate a dataset consisting of 100 traces to test the SMDM and RBPF automatically.
Figure 8 shows the precision, recall, and F-measure of the recognition results computed by the SMDM and AHMM-CTP.
In Figure 8, it is obvious that the performance of the SMDM is better when we have more observations.
To show detailed recognition results computed by the SMDM and RBPF, select two traces in the dataset.
The results show that the SMDM can recognize the destination efficiently no matter the intention changes or not, and it performs better than the AHMM-CTP in terms of precision, recall, and F-measure.