A Review on My Previous RL Work
nathanz.cloud/now/6
A Review on My Previous RL Work#
Over the past year or two, I have devoted much of my time to reinforcement learning agents. Having learned a great deal along the way, I am now finally able to pursue research of my own.
My recent focus has been DreamerV3, a world-model-based reinforcement learning agent. Building on it, I developed nine variants, each differing in design and in performance.
My strongest variant currently achieves 41.6% of the state of the art, an improvement of roughly 71% over the original DreamerV3.
| Method | Normalized return (% of 226) |
|---|---|
| ITC (2026), current SOTA | 7.09% ± 0.20 |
| Simulus (2025) | 6.59% |
| Improving Transformer World Models (2025) | 5.44% ± 0.25 |
| Mine: B2 | ~3.0% |
There is still a considerable gap to close, but my aim is clear: to develop future variants that surpass the current best.
The journey from student to independent researcher has been thrilling, and this is only the beginning. Every experiment brings me a step closer, and I could not be more excited about what lies ahead. The state of the art is in my sights, and I fully intend to claim it!