Forward-Backward Latent State Inference for Hidden Continuous-Time semi-Markov Chains

Authors: Nicolai Engelmann, Heinz Koeppl

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approaches in latent state inference scenarios in comparison to classical HSMM s.
Researcher Affiliation Academia Nicolai Engelmann Department of Electrical Engineering Technische Universitat Darmstadt 64289, Darmstadt nicolai.engelmann@tu-darmstadt.de Heinz Köppl Department of Electrical Engineering Department of Biology Technische Universitat Darmstadt 64289, Darmstadt heinz.koeppl@tu-darmstadt.de
Pseudocode Yes An algorithmic description and an explanation how φα and ψβ are obtained as by-products can be found in Appendix section C.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] In the Supplement
Open Datasets No We calculated several randomly generated hidden CTSMC s. The number of states has been given, but waiting time distributions and transition probabilities were drawn randomly. For each state, we drew a randomly parametrized waiting time distribution from a specified family of distributions (we allowed Gamma and Weibull). The random parameters, were drawn from Gamma distributions with per-parameter calibrated hyperparameters. The transition probabilities of the embedded Markov chains were sampled as sets of random categorical distributions while prohibiting self-transitions. The observation times were drawn from a point process with Gamma inter-arrival times with fixed shape and rate hyperparameters. The observation values were noisy observations of a function b : X R mapping the latent states to fixed values in a real observation space Y R with the Borel σ-algebra Y B. Only single points were allowed to be drawn, thus requiring density evaluations in the likelihood function υ. Therefore, Y (t) N(b(X(t)), d) with the standard deviation d as a hyperparameter. (This describes data generation, not the use or provision of a publicly available dataset.)
Dataset Splits No The paper describes generating synthetic data for simulations and comparisons, but does not explicitly mention train/validation/test splits for a dataset or a specific validation set.
Hardware Specification No The paper states that hardware information is in Appendix E ('Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] In Appendix E'), but Appendix E is not provided in the given text. Thus, no hardware specification is available in the provided text.
Software Dependencies No The paper discusses numerical methods like 'backward Euler method' and 'cubic spline interpolation' but does not specify versions of programming languages, libraries, or software packages used for implementation (e.g., Python 3.8, PyTorch 1.9).
Experiment Setup Yes For each time interval [t, t ) between observations, we used a backward Euler method to solve the possibly stiff equations in (7), (8) or (10) with a fixed number of steps calculated by N t t h + 1. [...] An exception is the evaluation of the adaptive step-size HSMM. As described in section 5, there the step size is initialized by a smallest value 10 4 and then the adaptive procedure generates any further grid positions. [...] With a step-size of h HSMM = 10 4 and h CTSMC = 10 3 (backward Euler step-size), no difference was visible anymore on any calculated trajectory.