This repository contains the code for the paper "Recursive Learning of Asymptotic Variational Objectives". It includes implementations of the OSIWAE algorithm, along with the associated proposal classes, as well as three key numerical simulations: run_ovsmc, run_slam, and run_growth. These simulations demonstrate the effectiveness of the OSIWAE method across different scenarios.

1. Introduction
2. Environment Setup
3. Key Files

This project demonstrates the implementation of the OSIWAE method, an extension of the Importance Weighted Autoencoder (IWAE) for state-space models (SSMs). The code also includes the associated proposal classes used by the OSIWAE algorithm to improve proposal adaptation.

Three numerical experiments are provided to reproduce the results from the paper:

• Multivariate Linear Gaussian (MVG) Model: Comparison with OVSMC and RML for parameter estimation in a multivariate Gaussian state-space model.
• SLAM: Application to simultaneous localization and mapping.
• Growth Model: Application to a highly non-linear state-space model.

The primary goal of these experiments is to showcase that OSIWAE consistently outperforms both OVSMC and RML in a variety of use cases, improving parameter learning and proposal adaptation.

To set up the environment for running the OSIWAE simulations, follow these steps:

1. Use the provided .yml file to create the environment:

   conda env create -f osiwae_env.yml

2. Activate the envirement:

   conda activate aistats
   
   

After setting up the environment, here’s an overview of the key files in the project:

• particle_algorithm.py: Class OSIWAE contains the main implementation of the OSIWAE algorithm.
• proposals.py: Includes the different proposal models used in the algorithm.
• models.py: Contains the three models we simulate: Multivariate Linear Gaussian, SLAM, and Growth Model.
• extra.py: Implements the particle sampling algorithms used within the simulations.

Additionally, the following scripts are used to run the numerical experiments:

• run_slam.py: For running the SLAM numerical experiment.
• run_mvg.py: For running the Multivariate Linear Gaussian (MVG) numerical experiment.
• run_growth.py: For running the Growth Model numerical experiment.