CatkinChen/nethack-hmm

Sticky-HDP-HMM for NetHack (Round 4/4)

A Sticky Hierarchical Dirichlet Process Hidden Markov Model trained on NetHack latent representations for learning skills and temporal dynamics.

Model Description

This is a Sticky-HDP-HMM model that learns discrete skills and temporal transitions in the latent space of a NetHack VAE. The model uses:

• Sticky self-transitions to encourage temporal persistence of skills
• Hierarchical Dirichlet Process for automatic skill discovery
• Normal-Inverse-Wishart priors for skill emission distributions

Model Details

• Model Type: Sticky Hierarchical Dirichlet Process Hidden Markov Model
• Framework: PyTorch with Variational Inference
• EM Round: 4 of 4
• Latent Dimensions: 96
• Maximum Skills: 40
• Base VAE: CatkinChen/nethack-vae

HMM Parameters

• Alpha (DP concentration): 5.0
• Kappa (sticky parameter): 1.0
• Gamma (top-level DP): 5.0

Usage

from train import load_hmm_from_huggingface
import torch

# Load the HMM
hmm, config = load_hmm_from_huggingface("CatkinChen/nethack-hmm")

# The HMM can be used with a VAE for skill-based generation

Training

This HMM was trained using Expectation-Maximization on VAE latent representations:

• E-step: Variational inference for posterior skill assignments
• M-step: VAE fine-tuning with HMM skill prior

Citation

If you use this model, please consider citing:

@misc{nethack-hmm,
  title={Sticky-HDP-HMM for NetHack Skill Learning},
  author={Xu Chen},
  year={2025},
  url={https://huggingface.co/CatkinChen/nethack-hmm}
}