Glen English VK1XX writes: > Has anyone tried to use a Neural net to control oven tmep, rather than > the ye olde PID ?
If you believe the marketing, that is why the Nest thermostat is connected to the cloud. > IE the algorithm learns from previous beheviour and successfully > predicts behaviour (or not). This part is (unsupervised) learning, not control. Depending on how much knowledge you bake into the neural network about the system, you either need to learn just the system parameters (say the PID constants) of a fixed system or figure out the system structure itself, the latter part is called system identification. Unless you expect the system to change over time it is usually (much) more practical to stop the learning at some point and just run the system with the best model available so far. Incidentally that is one reason why different hardware for neural network acceleration exists: the network for learning is usually much more complex and needs higher precision than the network running the extracted model. If you need to keep learning after the initial training (and if it has to be online learning rather than periodic offline re-training), the learning rate often has to be reduced significantly in order for the control to stay inside reasonable bounds. > I'm sure there are a few out there proficient with machine learning > algorithms. Neural networks are best applied to sparse, low-dimensional signals embedded in high-dimensional spaces. Oven temperature control doesn't fit that description, so more traditional methods are likely more efficient. > Might make a good masters thesis I bet. Maybe so, but there are probably sexier topics around for a master-to-be to pick up. > Given that oven control based on inputs and whatever is not random, > unlike say flicker etc. Randomness actually helps in many learning tasks. Regards, Achim. -- +<[Q+ Matrix-12 WAVE#46+305 Neuron microQkb Andromeda XTk Blofeld]>+ SD adaptation for Waldorf microQ V2.22R2: http://Synth.Stromeko.net/Downloads.html#WaldorfSDada _______________________________________________ time-nuts mailing list -- time-nuts@lists.febo.com To unsubscribe, go to http://lists.febo.com/mailman/listinfo/time-nuts_lists.febo.com and follow the instructions there.