Open oresthes opened 9 months ago
I must add that is seems to be an issue with the solver (or what is being passed to it) since the score method appears to be working correctly for the likelihood based learner
Hello @oresthes,
Actually optimizing for the llh of Hawkes processes with gradient descent is very hard due to the shape of the optimization curve (very flat near the optimum, very picky near the boundaries. Hence, the classical optimization algorithms (AGD, SVRG etc.) that rely on the gradient Lipschitz assumption have high chances to fail (see https://arxiv.org/abs/1807.03545).
You can try several hacks to make it work:
positive=True
in your penaltySee also https://github.com/X-DataInitiative/tick/issues/416#issuecomment-553423434
Hi!
I am using HawkesExpKern to infer parameters on a simulated process with known parameters. It is able to work ok(*) with least-squares as a goodness-of-fit measure but it struggles with likelihood. It errors out under most solvers and with svrg it fails to converge.
To replicate the process.
1) Simulate data
2) Infer using HawkesExpKern
3) Fit model
The response I am getting
If I repeat the same process using AGD as a solver it errors out as follows
What makes it even stranger is that I can find the maximum through brute force. This is the plot of the likelihood function (using the score method of the class). It converges a bit further away from the simulation parameters but it does exist.