You are assuming a linear relation between the dimension and the guessfactor, but that is true only for linear targeting. Won't that be a problem? -- ABC |

You are assuming a linear relation between the dimension and the guessfactor, but that is true only for linear targeting. Won't that be a problem? -- ABC Well, I wasn't proposing it as a new method all on it's own, I'm proposing combining this with DynamicClustering, and using this to refine the "cluster"'s data points for the current situation. So, yes, while it does assume linear relationships between dimensions and the guessfactor, and thus it won't get amazing results on it's own, it should serve well to refine the data from DynamicClustering somewhat. The point of it isn't to be perfect, but is to be better than just plain weighting the points in the cluster by their similarity to the current situation, by a technique repositioning the points to what they roughly "ought" to be in the current situation. I would estimate the lower bound of the performance of this, would be about same as the performance of the DynamicClustering, and the upper bound to be no better than the upper bound of the DynamicClustering performance, however I think it may be able to give more refined results when there are relatively few available data points.The more I think about it though, it may be interesting to toss the DynamicClustering and just use recent logs instead of near neighbors, and replace the linear regression with a separate neural network for each cluster (after all neural networks are good for being nonlinear function approximators). For that to work, I would need to toss out old data and only keep a certain length of logs, due to not being able to optimize it like a KDTree, but it could be an interesting hybrid of neural, clustering, and log based. Actually, I think I may give that a try some day... -- Rednaxela |