So far my thesis topic has been a big learning curve for me. I had never really thought of trying my hand at game design, which is why I was slightly hesitant when I came to understand that the game engine I would be using for experimentation was somewhat “a work in progress”.  Throughout the uni break I was working almost diligently on modification of the game mechanics to something I was happy to start working with. It is a pleasure to say that I am almost there.

I am happy to report that I am pretty much to schedule with what I had intend to have ready by second semester. I have namely been experimenting with different  methods of motion planning and am surprised as how something so simple can drastically change the feel of game play. The two methods I have mainly been looking at are;

1. Delauny Triangularization using equidistant points between the level boundaries
2. Random Roadmap methods which employ generating a number of random points across the level and running Delauny triangularization across them as to create suitable paths for the AI to explore.

While the first method is quite simplistic and reliable, the paths the AI take can be reasonably limited in the exploration space. The second methods is interesting as every game is slightly different, and with the monotony of testing it is nice to have the playing ground mixed up; but whether this is suitable for an actual, solid implementation, I’m not sure. What I am doing at the moment is working on constructing my own paths about the level, “hand drawn” if you will.

As far as tracking goes I have implemented a shared particle filter. Every guard contributes to the maintaining and updating of the single distribution. I have found, while being a very simplistic approach, that it is unsuitable. During the resampling portion of the particle filter algorithm, a single focal point tends to emerge when the distribution contains less than about 500 particles. This becomes incredibly boring to play with as every guard then heads towards this single point, converging uselessly – especially when the centre of the mass is chosen. The guards paths then tend to overlap more than my liking, and this is not the behavior I would like to see emerging.

So far I have not experimented to much with planning algorithms, but the night is still young, as it were. So far I have only looked at choosing points at random from the cloud or sampling them in order to find the mean of the distribution. As above, sampling the mean of the single distribution makes the game play quite boring. What I would like to try is taking a handful of particles, randomly, and finding their mean as to determine a psuedo-mean. I am hoping this method will provide what I originally sort after and reduce the overlapping of search paths.