Haven't updated this in a while. In this version I changed the weighting system for the neurons and I also switch the
NN to use a single row of neurons rather than a grid.
Haven't updated this in a while. In this version I changed the weighting system for the neurons and I also switch the NN to use a single row of neurons rather than a grid.
Rolled Back to 0.3.6 + changes. This version marks a big change. I have now fully implemented the Kohonen network; I had originally cut a few features out for execution speed reasons. now I believe that I have worked out most of the issues, it's a bit slower than previous versions but not too much. I still have to tune it, but the network should now provide much better classifications of inputs. Rating: ?
I havn't updated this page in a while. All versions between 0.3.3 and 0.4.0 have been mostly tweaks to various weights and parameters in the Kohonen Network. In this release I am experimenting with several new dimensions in the classification vectors for the network that drives the gun. I also tweaked the gun's network to allow more nodes to be created, which should result in more specific classifications. Rating: 2031
I decided to switch my focus to the gun for a little while, as it is significantly weaker than the movement. The gun in this version appears to be stronger than the previous version by a good bit, but the Rumble will tell for sure. Rating: ?
fixed a few really stupid bugs in the surfing code, including a division by zero that sometimes happened when smoothing the data. Rating: 2030
fixed a null pointer exception that occurred during close fighting. Rating: 2029
Added dynamic decisions for distancing based on enemy hit rates. I've been avoiding this because it seems like a double-edged sword; it may help me avoid an enemy's gun, but now my gun will have to deal with different distances as well. Rating: pulled
Rolled back to v. 0.2.5 and corrected a long-standing bug in the internal workings of the neural nets. Rating: 2030
I tried to increase the learning speed of the gun's network. Don't know how successful it will be in the Rumble though. Rating: 2002 (yuck)
Tweaked the wall proximity dimesion in both the gun and the surfing. Made some changes to the classification vectors for the gun. Rating: 2029
I've been visiting Europe for the past three weeks and had no access to Robocode, so this version has only some small tweaks to the classification vectors while i re-learn my code. Rating: 2031
Changed the way that input vectors are blended with the node weighting vectors. Rating: 2027
Changed the proximity formula to use normalized euclidean distance. Also this version is using a dynamically growing SOM, which is significantly faster. Rating: 2026
Fixed a wave bug I introduced accidentally in 0.2.1. Rating: 2026
Some tweaks to the surfing weighting vectors as well as some performance optimization. If nothing else this version is faster to execute. Rating: 2022
Now using SOM for surfing as well. It needs to be tuned more, right now the surfing network is using the same weighting vectors as the gun's network. Rating: 2026
Initial release. Using Self Organizing Map for targeting and PowerHouse's wavesurfing. Rating: 2030