Hey guys, so I have been working on actuator-net which introduces the concept of learning model parameters of an actuator so that it can be later used in simulation to bridge the simulation to reality gap in robotics. The idea is to collect data from the hardware in form of position errors, velocity feedback and torque estimate from the torque sensor. The input to the model will be a history of position errors, velocity feedback and the model should output torque estimates for which I am employing a neural network of which has 2 hidden layers first with 8 neurons and second with 4 networks which connects to the output layer of 1 neuron, I have done all the preprocessing and everything but I cannot get the model to not overfit it starts overfitting in first few epochs itself, what to do?
The usual fix for overfitting is adding regularization.
This may also take the form of using dropout.
I tried that but it’s not working, I put a dropout layer of 0.3 but it didn’t work.