Learning the conditional Law. Filtering and Prediction

Code for the paper Learning the conditional law: signatures and conditional GANs in filtering and prediction of diffusion processes, accepted at CDC 2022.

Abstract:

   We consider the filtering and prediction problem for a diffusion process. The signal and observation are modeled by stochastic differential
   equations (SDEs) driven by Wiener processes. In classical estimation theory, measure-valued stochastic partial differential equations (SPDEs)
   are derived for the filtering and prediction measures. These equations can be hard to solve numerically. We provide an approximation algorithm
   using conditional generative adversarial networks (GANs) and signatures, an object from rough path theory. The signature of a sufficiently
   smooth path determines the path completely. In some cases, GANs based on signatures have been shown to efficiently approximate the law of a 
   stochastic process. In this paper we extend this method to approximate the prediction measure conditional to noisy observation. We use 
   controlled differential equations (CDEs) as universal approximators to propose an estimator for the conditional and prediction law. We show
   well-posedness in providing a rigorous mathematical framework. Numerical results show the efficiency of our algorithm.


@misc{https://doi.org/10.48550/arxiv.2204.00611,
  doi = {10.48550/ARXIV.2204.00611},
  url = {https://arxiv.org/abs/2204.00611},
  author = {Germ, Fabian and Sabate-Vidales, Marc},
  keywords = {Machine Learning (stat.ML), Machine Learning (cs.LG), Optimization and Control (math.OC), FOS: Computer and information sciences, FOS: Computer and information sciences, FOS: Mathematics, FOS: Mathematics},
  title = {Learning the conditional law: signatures and conditional GANs in filtering and prediction of diffusion processes},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license} 
}

The problem

Consider the signal-observation system where the coefficients are assumed to be regular enough (see Assumption 2.1 of the paper.

The goal of filtering and prediction is to derive and analyze the quantities

for any continuous \varphi, i.e. we train a generator that is able to sample from the conditional law, and we approximate the above quantities using averages and law of large numbers.

The method

We use the Conditional Sig-Wasserstein GAN, using the following Neural Differential Equations for the Generator. The generator is divided in two, the past and the future. Fix the present at time s. Then the generator is given by the two models below.

The generator is well posed. (3) is can be interpreted as the process carrying the information known up until the present s. (4) can be interpreted as the generation of possible future scenarios, given the information known at the present s. More specifically, we prove the form it should actually have in Lemma 4.2.

The code

Relevant scripts:

lib/nde.py parametrizes the Neural CDE and Neural ODE of the generator. They are solved using torchcde and torchdiffeq respectively.
lib/sigcwgan.py implements the sigcwgan class and its fitting methods.
lib/data.py creates a synthetic dataset from the SDE using Euler method. To be changed for new SDEs. This script also implements the Kalman filter in the linear case.

Requirements: to set up the environment:

conda env create -f environment.yml

Training:

usage: train.py [-h] [--base_dir BASE_DIR] [--device DEVICE] [--use_cuda] [--seed SEED]
                [--num_epochs NUM_EPOCHS] [--depth DEPTH] [--T T] [--n_steps N_STEPS]
                [--window_length WINDOW_LENGTH]

optional arguments:
  -h, --help            show this help message and exit
  --base_dir BASE_DIR   base dir for results
  --device DEVICE       gpu id if cuda is available
  --use_cuda
  --seed SEED
  --num_epochs NUM_EPOCHS
  --depth DEPTH         depth of signature
  --T T
  --n_steps N_STEPS     number of steps in time discretisation
  --window_length WINDOW_LENGTH
                        window length used in log-ODE method

Evaluation. Right now there is only evaluation in the linear case, using the Kalman filter.

usage: evaluate.py [-h] [--base_dir BASE_DIR] [--device DEVICE] [--use_cuda] [--seed SEED] [--depth DEPTH] [--T T] [--n_steps N_STEPS] [--window_length WINDOW_LENGTH]

optional arguments:
  -h, --help            show this help message and exit
  --base_dir BASE_DIR   base dir for results
  --device DEVICE       gpu id if cuda is available
  --use_cuda
  --seed SEED
  --depth DEPTH         depth of signature
  --T T
  --n_steps N_STEPS     number of steps in time discretisation
  --window_length WINDOW_LENGTH
                        window length used in log-ODE method

Plots are saved in the value given to args.base_dir, which is numerical_results by default.

TODOs

Code the splitting-up to solve the Zakai eq for evaluation
- PDE solver (FEM or DL in higher dims)
- Sampling from an unnormalised density (MALA, MCMC, Langevin? --> need the gradient of the sol of the zakai equation, ugh)
Solve the forward kolmogorov eq for transition prob
Jumps in the diffusion proces --> the generator will need to be changed. There is always interpolation in the driver of the CDE/rough neural differential equation, so eventually everything is always of bounded variation.... ODE-RNN method for the generator? What about the discriminator?
Control problems...

msabvid/SigFiltering

Learning the conditional Law. Filtering and Prediction

The problem

The method

The code

TODOs