Given general linear differential equation in matrix vector multiplication form [dx(t)/dt = Ax(t)+Cu(t)], I tried to estimate the coefficients of the linear differential equation. Given the parameters of the model, which are coefficients (A,C, both are matrix) and stimulus values (u(t)), and the initial state of the system x(0), the response or solution x(t) can be determine numerically. This is the forward problem.
However, in this project, I solve inverse problem through experiment results. The inverse problem is the problem of finding the parameters in the model from a sequence of experiments. Specially we want to determine A and C given a sequence of experiments where we input different functions u(t) and measure the output x(t). And since x depends on the parameters, I can write the solution as x(t; A;C). When I conduct at experiment, measurements are always contaminated by noise and samples are only taken at a discrete set of point say {t_1, ..., t_k}. So instead of being able to measure x directly we instead measure: y^i(t_k) = x^i(t_k) + e, where e is Gaussian random variable. So I try to estimate coefficients through experiments by minimizing the squared error.
Matlab. I coded the optimization algorithm using Matlab.
Comparing the mismatch between original combined model parameter and estimated model parameter (err ), I can say the followings:
- Choosing appropriate input function is important.
- Our optimization algorithm is sensitive to noise a little bit. We can see this from the error calculations of cases. Each time our estimated means square error (mse) is lower than original mse. This shows that our model tried to mask the effect of noise as well. However, difference is not that much. I can say that sensitivity of the algorithm to my noise was negligible.
- For more experiment cases (10 experiments) our mismatch is little bit higher than that of in less experiment cases (5 experiments). I think, this is because our model become more sensitive to noise . We can see this if we compare the estimated and original mse. The reason for the model being sensitive to the noise of the data is that when we use more training data (more experiments) we may end up with overfitting to the noise.