Thanks very much, Matthew. You indeed have made good progress! Do you think the model is currently attempting to infer future tilt from past temperature?
>> "and the model params it returned did not include any correlation between Tilt_06 and any temperature inputs" Do you mean the HTM swarming did not discover any correlation between tilt and previous temperature? If so that is a problem... >> "but the model is not getting trained enough" This seems to be what we observed as well. The HTM seemed to be constantly surprised by the new data and anomaly score did not go down. If we switched off learning, the predictions almost immediately became very poor meaning (I think) the model was constantly having to learn each new sequence as it came along and was not generalising the data. It never seemed to gain a real understanding however long it was trained for. If you just consider a single stream, e.g., one of the temperatures, it looks superficially similar to the hotgym data but actually there are profound differences. It depends on the weather which is largely unpredictable so lacks the repetitiveness of the hotgym data. Thus, we cannot expect it to learn with just one stream, IMO. The HTM needs to master the cross-correlations. On Tue, Sep 30, 2014 at 3:27 PM, Matthew Taylor <[email protected]> wrote: > > On Tue, Sep 30, 2014 at 3:47 AM, John Blackburn < > [email protected]> wrote: > >> If I set the predicted field as "tilt" will it refrain from trying to >> predict the temperature also? I want to avoid it trying to predict >> temperature. Basically I want to shut down some of the prediction >> combinations in favour of the useful ones. I could imagine doing this >> manually (but don't know how to do it) but the best would be if the HTM can >> figure out the useful combinations itself. I think you are saying swarming >> does this? I will see if it works. > > > John, > > I started working on your problem last week [1]. Yes, if the predicted > field is one of the tilt field, swarming will not optimize for anything > except prediction accuracy on the specified tilt field. I ran a limited > swarm on Tilt_06 predictions, and the model params it returned did not > include any correlation between Tilt_06 and any temperature inputs. I'm > currently using model params that are tuned for anomaly detection, and I'm > not using any temperature input. I may or may not add back temperature > encoders to the params depending on how the model learns. > > The next step is to train a model on the first two years of data and > persist it. Then I want to de-serialized the model and push through the > last year of data, plotting all the anomalies. I've got the anomalies > plotting now ("./run_anomaly.py --plot"), but the model is not getting > trained enough. And I'll probably need to get some help tuning the params. > > But I did make some progress last week. > > [1] https://github.com/nupic-community/bridge-tilt > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta >
