Add support for Stable-Baselines3, a PyTorch-based library providing reliable, well-tested implementations of popular deep reinforcement learning algorithms with a simple and consistent API for research and training.

docs/scientific-packages.md

@@ -50,6 +50,7 @@
 - **scikit-learn** - Machine learning algorithms, preprocessing, and model selection
 - **scikit-survival** - Survival analysis and time-to-event modeling with censored data. Built on scikit-learn, provides Cox proportional hazards models (CoxPHSurvivalAnalysis, CoxnetSurvivalAnalysis with elastic net regularization), ensemble methods (Random Survival Forests, Gradient Boosting), Survival Support Vector Machines (linear and kernel), non-parametric estimators (Kaplan-Meier, Nelson-Aalen), competing risks analysis, and specialized evaluation metrics (concordance index, time-dependent AUC, Brier score). Handles right-censored data, integrates with scikit-learn pipelines, and supports feature selection and hyperparameter tuning via cross-validation
 - **SHAP** - Model interpretability and explainability using Shapley values from game theory. Provides unified approach to explain any ML model with TreeExplainer (fast exact explanations for XGBoost/LightGBM/Random Forest), DeepExplainer (TensorFlow/PyTorch neural networks), KernelExplainer (model-agnostic), and LinearExplainer. Includes comprehensive visualizations (waterfall plots for individual predictions, beeswarm plots for global importance, scatter plots for feature relationships, bar/force/heatmap plots), supports model debugging, fairness analysis, feature engineering guidance, and production deployment
+- **Stable Baselines3** - PyTorch-based reinforcement learning library providing reliable implementations of RL algorithms (PPO, SAC, DQN, TD3, DDPG, A2C, HER, RecurrentPPO). Use this skill for training RL agents on standard or custom Gymnasium environments, implementing callbacks for monitoring and control, using vectorized environments for parallel training, creating custom environments with proper Gymnasium API implementation, and integrating with deep RL workflows. Includes comprehensive training templates, evaluation utilities, algorithm selection guidance (on-policy vs off-policy, continuous vs discrete actions), support for multi-input policies (dict observations), goal-conditioned learning with HER, and integration with TensorBoard for experiment tracking
 - **statsmodels** - Statistical modeling and econometrics (OLS, GLM, logit/probit, ARIMA, time series forecasting, hypothesis testing, diagnostics)
 - **Torch Geometric** - Graph Neural Networks for molecular and geometric data
 - **Transformers** - State-of-the-art machine learning models for NLP, computer vision, audio, and multimodal tasks. Provides 1M+ pre-trained models accessible via pipelines (text-classification, NER, QA, summarization, translation, text-generation, image-classification, object-detection, ASR, VQA), comprehensive training via Trainer API with distributed training and mixed precision, flexible text generation with multiple decoding strategies (greedy, beam search, sampling), and Auto classes for automatic architecture selection (BERT, GPT, T5, ViT, BART, etc.)