Answers below... On Thu, Feb 5, 2015 at 5:31 PM, Matsushita,Toshiyuki <MKI USA> <[email protected]> wrote: > Hello, > Now, I have just started to explore predictive possibilities of NuPIC. > Although I believe these are very basic questions for predictive functions of > NuPIC that you may be already understanding, I would appreciate if someone > could advise or give answers to the following questions regarding the sample > of Hot Gym Prediction and CPU sample. > > ------------------------------------------------------------------ > <1> Hot Gym Prediction > ------------------------------------------------------------------ > > Q1 : Is it possible to lead to answers at one time as predicted data, > anomalyScore and anomalyLikelihood by feeding different data streams acquired > from different data sources such like GYM1, GYM2 or GYM3? > > [Input data] > GymID Date Consumption > GYM1 2/5/2015 0:00:00 21.2 > GYM2 2/5/2015 0:00:00 12.3 > GYM3 2/5/2015 0:00:00 31.5 > GYM1 2/5/2015 1:00:00 16.4 > GYM2 2/5/2015 2:00:00 11.8 > GYM3 2/5/2015 3:00:00 30.5 > : : : > : : : > GYM1 2/5/2015 23:00:00 11.2 > GYM2 2/5/2015 23:00:00 2.3 > GYM3 2/5/2015 23:00:00 21.5 > > > [Desirable Output] (Prediction) > GymID Date Consumption anomalyScore > anomalyLikelihood > GYM1 2/6/2015 1:00:00 16.3 0 > 0.5 > GYM2 2/6/2015 1:00:00 11.5 0 > 0.3 > GYM3 2/6/2015 1:00:00 29.1 0 > 0.2
Assuming the the gyms are independent, meaning the energy consumption at one gym is completely independent of the energy consumption at other gyms, you should create a model for each gym instead of trying to send all the gyms' data into one model. This means that each gym will have a model created for it, and each one gets its own data passed into its model. Then each model will learn the patterns within its gym and make predictions only based on that gym's energy consumption. > Q2 : In addition to above, how I could write scripts with JSON to execute > swarm in case of the model_params which predicts analytics results above? In this case, the models parameters for each gym may be close enough that you won't need to swarm for each gym. Just use the same model parameters for each model you create. As the models get passed data, they will learn only the patterns in the data for the gym they represent. If you want to swarm once for each gym, the swarm.py script can be modified to generate a swarm description for each gym. I started putting together a "Many Hot Gyms" tutorial last year, but haven't gotten around to completing it. But you might find the code useful at https://github.com/rhyolight/nupic/blob/many-hot-gyms-prediction/examples/opf/clients/hotgym/prediction/many_gyms/ (particularly the swarm script at https://github.com/rhyolight/nupic/blob/many-hot-gyms-prediction/examples/opf/clients/hotgym/prediction/many_gyms/swarm_helper.py). > ------------------------------------------------------------------ > <2> CPU Sample > ------------------------------------------------------------------ > > Q1 : Is it possible to lead to answers at one time which are categorized by > different parameters by feeding data streams obtaining these parameters such > like CPU(%), Memory(GB)and DISK_USAGE(GB) as shown below? > > [Input data] > CPU(%) Memory(GB) DISK_USAGE(GB) > 12.3 75.6 250.4 > 15.6 68.5 251.3 > 13.7 71.6 251.8 > > [Desirable Output] (Prediction) > CPU(%) Memory(GB) DISK_USAGE(GB) > 14.8 69.7 252.1 While a model can accept many fields of data at once (representing factors that might affect the predicted field), it can only output a prediction for one field. In order to get predictions for 3 fields like you've defined above, you'll need to create 3 models. > Q2 : In addition, how I could write scripts with JSON to execute swarm in > case of the model_params which predicts analytics results above? See my answer to your other Q2 above. Regards, --------- Matt Taylor OS Community Flag-Bearer Numenta
