|
Oooh, interesting. Does this do it on the level of how buffers are merged, or on the level of changing the actual context of the buffers? If the former, this could be extremely similar to my dynamically weighted CrowdTargeting that I'm currently trying to improve :) -- Rednaxela |
|
Oooh, interesting. Does this do it on the level of how buffers are merged, or on the level of changing the actual context of the buffers? If the former, this could be extremely similar to my dynamically weighted CrowdTargeting that I'm currently trying to improve :) -- Rednaxela Well, first attempts don't seem to do much. Basically what I'm doing is, when logging hits, seeing which sets of bins (extracted from the segments) predicted that hit correctly. The ones that predict right get weighted higher. The important part is weighting *that* set of bins at *that* location in the buffer, so that a different location in the same buffer doesn't benefit from the weighting. Applying this to DynamicClustering would be a bit harder though. Perhaps keeping track of the shots that make up a wave, and increasing the 'importance' of the ones that result in a successful hit? However, one problem I now see with this is a constant shuffling of buffers because *if* it gets a hit, it is due to the fact that currently high-weighted buffers didn't predict it, whether or not they happen to be the best. Maybe it would be better to only use this on 'unbiased' buffer additions, ie. bullet-hit-bullet hits. Giving that a try now =) -- Skilgannon |
After I get some tests done I'll write up a full description. Including whether it helps or not =) -- Skilgannon
Oooh, interesting. Does this do it on the level of how buffers are merged, or on the level of changing the actual context of the buffers? If the former, this could be extremely similar to my dynamically weighted CrowdTargeting that I'm currently trying to improve :) -- Rednaxela
Well, first attempts don't seem to do much. Basically what I'm doing is, when logging hits, seeing which sets of bins (extracted from the segments) predicted that hit correctly. The ones that predict right get weighted higher. The important part is weighting *that* set of bins at *that* location in the buffer, so that a different location in the same buffer doesn't benefit from the weighting. Applying this to DynamicClustering would be a bit harder though. Perhaps keeping track of the shots that make up a wave, and increasing the 'importance' of the ones that result in a successful hit?
However, one problem I now see with this is a constant shuffling of buffers because *if* it gets a hit, it is due to the fact that currently high-weighted buffers didn't predict it, whether or not they happen to be the best. Maybe it would be better to only use this on 'unbiased' buffer additions, ie. bullet-hit-bullet hits. Giving that a try now =) -- Skilgannon
![[Home]](/images/RoboWiki.png)