Replica Conditional Sequential Monte Carlo
Authors: Alex Shestopaloff, Arnaud Doucet
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We propose here replica c SMC where we build SMC proposals for one replica using information from the entire observation sequence by conditioning on the states of the other replicas. This approach is easily parallelizable and we demonstrate its excellent empirical performance when compared to the standard iterated c SMC scheme at fixed computational complexity. [...] In Section 4, we demonstrate the methodology on a linear Gaussian model, two non-Gaussian state space models from (Shestopaloff & Neal, 2018) as well as the Lorenz-96 model from (Heng et al., 2017). |
| Researcher Affiliation | Academia | 1School of Mathematics, University of Edinburgh, Edinburgh, UK 2The Alan Turing Institute, London, UK 3Department of Statistics, University of Oxford, Oxford, UK. |
| Pseudocode | Yes | Algorithm 1 Iterated c SMC kernel K (x1:T , x 1:T ) c SMC step. [...] Algorithm 2 Replica c SMC update For k = 1, . . . , K |
| Open Source Code | Yes | The code to reproduce all the results is publicly available at https://github.com/ ayshestopaloff/replicacsmc. |
| Open Datasets | No | The paper uses synthetic data generated from described models (Linear Gaussian, Poisson-Gaussian, Lorenz-96 models). It explicitly states: 'We generate a sequence from this model to use for our experiments.' and 'We generate one sequence of observations from each model.' No existing publicly available datasets are utilized and referenced with concrete access information. |
| Dataset Splits | No | The paper evaluates a sampling algorithm on generated time series data. It does not describe traditional training, validation, or test dataset splits as would be relevant for a predictive model. The evaluation focuses on sampler properties like autocorrelation time. |
| Hardware Specification | Yes | To do our computations, we used MATLAB on an OS X system, running on an Intel Core i5 1.3 GHz CPU. |
| Software Dependencies | No | The paper states: 'To do our computations, we used MATLAB on an OS X system'. While 'MATLAB' is named, no specific version number for MATLAB or any other software libraries/dependencies are provided for reproducibility. |
| Experiment Setup | Yes | We do 10 replica c SMC runs with 100 particles and 2 replicas for 25, 000 iterations. [...] We do 5 runs of each sampler. Both samplers use 100 particles and we do a total of 5, 000 iterations per run. [...] We use replica c SMC with 5 replicas, updating one replica conditional on the other. [...] We set the number of particles to 200. We do a total of 5 runs of the sampler with 5, 000 iterations. [...] Both replica c SMC and iterated c SMC updates use 100 particles. [...] We run replica c SMC with 200 particles for 30, 000 iterations [...] and compare to standard iterated c SMC with 600 particles, which we also run for 30, 000 iterations. |