trtpi
Code underlying the publication: Trust-Region Twisted Policy Improvement
Description
TRT-SMC: Trust-Region Twisted Sequential Monte-Carlo — JAX implementation accompanying the ICML 2025 paper "Trust-Region Twisted Policy Improvement". Combines MCTS-inspired design choices (trust-region constrained action sampling, explicit terminal handling, improved policy/value targets) with particle filter planning to construct a GPU-efficient SMC policy improvement operator. Outperforms baseline MCTS and SMC planners in runtime and sample-efficiency across discrete and continuous RL domains. The smz module exposes a modular SMC planner accepting custom Proposal, Transition, and Target implementations, usable as a drop-in policy replacement compatible with jax.jit and jax.vmap.
- MIT
Reference papers
Mentions
- 1.Author(s): de Vries, Joery Ariën, Lawrence, Neil David, Dai, ZhenwenPublished in 202610.48550/arxiv.2610.03361
- 2.Author(s): Oren, Yaniv, Vadocz, Viliam, de Vries, Joery A., Böhmer, Wendelin, Spaan, Matthijs T. J., Baier, HendrikPublished in 202610.48550/arxiv.2605.08982
- 3.Author(s): Oren, Yaniv, Vadocz, Viliam, Zabka, Wiktor, Evers, Thomas, Robine, Jan, Böhmer, Wendelin, Spaan, Matthijs T. J., White, Martha, Baier, Hendrik, Yu, FenghuiPublished in 202610.48550/arxiv.2610.01566
- 4.Author(s): Oren, Yaniv, de Vries, Joery A., van der Vaart, Pascal R., Spaan, Matthijs T. J., Böhmer, WendelinPublished in 202510.48550/arxiv.2511.14220