Add model_card.md

model_card.md

@@ -0,0 +1,89 @@
+# UGTC Model Card
+
+## Model Description
+
+**Name:** UGTC (Uncertainty-Gated Temporal Credit)
+**Type:** Reinforcement Learning Algorithm Component (Advantage Estimator)
+**Author:** Yağız Ekrem Dalar
+**License:** MIT
+**Paper:** https://doi.org/10.5281/zenodo.19715116
+
+### What is UGTC?
+
+UGTC is a **plug-in advantage estimator** for actor-critic reinforcement learning algorithms.
+It is not a standalone model but a module that replaces the standard advantage computation
+in any actor-critic algorithm (PPO, TD3, SAC, DDPG, etc.).
+
+### How it Works
+
+UGTC maintains two value critics with different GAE λ values:
+
+1. **Fast critic** (single MLP, λ=0.80): Low variance, higher bias
+2. **Slow ensemble** (M=3 MLPs, λ=0.99): Lower bias, higher variance
+
+A sigmoid gate `u(s) = σ(-β·(σ̂(s) - 1))` blends their advantage estimates:
+
+```
+A^UGTC_t = u(sₜ)·A^slow_t + (1-u(sₜ))·A^fast_t
+```
+
+where `σ̂(s)` is the EMA-normalized ensemble disagreement.
+
+## Intended Use
+
+- Research on advantage estimation in deep RL
+- Integration into existing actor-critic implementations
+- Educational exploration of uncertainty-based credit assignment
+
+## Fixed Hyperparameters
+
+| Parameter | Value | Description |
+|-----------|-------|-------------|
+| λ_fast | 0.80 | Fast critic GAE lambda |
+| λ_slow | 0.99 | Slow ensemble GAE lambda |
+| M | 3 | Ensemble size |
+| β | 5.0 | Gate temperature |
+| EMA momentum | 0.99 | Running uncertainty normalization |
+
+These are fixed across **all benchmarks** — no per-task tuning.
+
+## Training and Evaluation
+
+UGTC has been integrated with PPO, TD3, and SAC and evaluated on:
+- Procgen 16-game benchmark (hard mode, 25M steps)
+- MetaWorld ML45 (multitask)
+- DeepMind Control Suite (DMC)
+- ManiSkill (manipulation multitask)
+- MuJoCo: Hopper-v4, Ant-v5
+- Image-based: CarRacing-v3
+
+For specific results, see the paper: https://doi.org/10.5281/zenodo.19715116
+
+## Limitations and Assumptions
+
+- UGTC-DDPG is an **implementation extension** — not directly evaluated in the paper
+- Performance gains depend on the specific environment and algorithm backbone
+- The UGTC critics introduce a small parameter overhead (~3 MLP value heads)
+- EMA normalization requires a warmup period for accurate σ_EMA estimation
+- Performance may vary from paper results due to hardware, library versions, and random seed
+
+## Ethical Considerations
+
+UGTC is a research algorithm for reinforcement learning. It does not involve:
+- Personal data
+- Decision-making in high-stakes autonomous systems
+- Generative content
+
+## Publication
+
+> **UGTC: Uncertainty-Gated Temporal Credit**
+> Yağız Ekrem Dalar
+> *Accepted — Ulysseus Young Explorers in Science (UYES) Journal*
+> Preprint DOI: [10.5281/zenodo.19715116](https://doi.org/10.5281/zenodo.19715116)
+> Journal DOI: Forthcoming
+
+## Links
+
+- GitHub: https://github.com/ethosoftai/ugtc
+- Documentation: https://ethosoftai.github.io/ugtc
+- Demo: https://huggingface.co/spaces/Ethosoft/ugtc