Proposal 1

To investigate how the DG system can be used to automatically generate similar" motions to what the user supplies as input. The idea is for the user to give to the DG one or more examples of a style of motion (limp, happy walk etc.), and for the system to automatically generate variations on that style (severe limp, not-so-happy walk).

The above statement corresponds to the DG doing a search in the n-dimensional space of the input control parameters to find points which will yield "similar" animations. Thus the probability of the system finding points which represent "interesting" motions is quite small. Although dispersion helps to distribute the points considered throughout the search space, it would help further if we could bound the search space in some way. One method is to have the user input more than one motions, where each one represents a different degree of the same style (happy walk, sad walk). These parameters partition the space further, helping in concentrating our efforts in this smaller region. Further, a knowledge-based optimization function can be built in, which can further help to identify "interesting" solutions.

Proposal2

To investigate distance metrics which will ascertain how similar two animations are in their stylistic content. For example, I would like a distance metric which would tell me that a bad limp and slight limp are "close" to each other, while a happy walk is "far".

One way to go about constructing these distance metrics is that instead of working for a long time (overnight), and then displaying the results to the user, the DG can work for a short period of time (1 hr ?), and present the user with a small number (3-4 ?) of solutions which the user can "grade". At this point the algorithm will have the original input from the user as well as "grades" for the solutions. These can be used to tune the distance metric hard-coded in the program. With the improved distance metrics in hand, the DG can now work overnight and the solutions it generates should match more closely to the intentions of the user. I believe that this approach is better than user-guided exploration in that the feedback from the user is not being used to limit the search, but to compare the results obtained. Thereby ensuring that only those results are shown to the user which are "close" to the input.