R gbm: Lower shrinkage gives worse results? I've read here and on other sites that when using GBM in R lowering shrinkage gives better results. Yet in my case this is clearly not the case. 0.1 is better than 0.01 with same amount of trees. Even with more trees 0.01 gives worse results.
Is that generalization just plain wrong or what could case this?
 A: More shrinkage (higher learning rate) gives lower long-term accuracy.  It has higher rates of error.  It has high initial convergence rates, but after it reaches steady-state, the level is relatively high.
Lower shrinkage (lower learning rate) gives higher long-term accuracy.  It has lower rates of error.  In the beginning it has lower rates of learning, but after it has reached steady-state, the level is typically lower than for the alternative rates.
So if you jump into the learning process early, you would get that the "high rate" has lower error and the "low rate" has high error.  If you waited until the process was much farther along, then you could see a more expected outcome.
Personally, I have wondered about rate-switching, and rate-selective pruning, but have not had a chance to experiment there.  
Best of luck.
EDIT
If you are overfitting a really simple system, like the first 5 trees do the job, then you can also get this phenomena.
Also, be sure your number is learning. If it is 1-learning then you get contrary behavior.
