Deep Reinforcement Learning Presentations

This repository contains lecture materials from two presentations I delivered at the Faculty of Informatics, Masaryk University, as part of the IV125 Formela Lab Seminar focused on advanced reinforcement learning topics.

Presentations

1. Discovered Policy Optimisation (DPO)

presented on 2025-03-07

• Mirror Learning Framework: PPO paradox, drift functions, theoretical foundation
• Meta-Learning Discovery: Evolution strategies, automatic algorithm discovery
• DPO Formula: From open to closed-form formula, outperforms PPO

2. Group Relative Policy Optimisation (GRPO)

presented on 2025-05-16

• GRPO vs PPO: Value model elimination, group-relative advantage
• DeepSeekMath: Base model, continual pre-training, SFT, GRPO
• DeepSeek-R1: Pure RL, GPT-o1 performance, mathematical reasoning

---

Disclaimer: Some figures are reproduced from original research papers for educational purposes. All sources are properly cited.