To all:
There has been talk on previous posts discussing the value of annealing. I have always advocated using it and could only give "fine tuning" as a reason. Well, I just saw something happen that, to me, gives good reason for using annealing, and I thought I should make everyone aware of it.
I was doing my weekend net training and was watching a particular net closely because I already had two consecutive buy signals from daily back prop training from last week and three is my signal to take a position. It trained on back prop for sometime and produced the third buy signal with a positive outlook of 1.0 and a negative outlook of 0. Since the weekend is when I "update" my nets completely , I continued training and allowed it to go into annealing. After training over 25000 cycles, the green volatility bars at the end of the chart had not changed. Also, I noticed that the "Performance with Neurostock" number was not changing, a very good sign that the net is no longer learning anything. So I stopped the training and hit Predict and got a Hold signal with a positive outlook of 1.0 and negative outlook of 0.6. Now I'm doing a double take because I was expecting the buy signal. Since most of the time annealing widens the line in the stats window a touch and had done so this time, I decided to tighten it up with some back prop. As soon as I started the training again, not only did the last green bar disappear but so did previous two. I trained on back prop for over 2000 cycles and the prediction did not change. Without doing the annealing training I would have probably taken a position in this stock on Monday. Now I won't touch it. Time will tell if no action was the action to take.
Sorry to be so long winded, but I had nothing better to do.
Jay |
