Reproduce “DESTA: A Framework for Safe Reinforcement Learning with Markov Games of Intervention” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
David Mguni, Usman Islam, Yaqi Sun, Xiuling Zhang, Joel Jennings, Aivar Sootla, Changmin Yu, Ziyan Wang and 2 more
DESTA separates a reward-seeking task policy from a learned safety policy that selectively takes control when proposed actions risk safety costs. The safety policy learns an action and when to intervene, accounting for future costs. On the one-way Shortest Safe Route task, DESTA-SAC learns the shortest safe route, reaching the maximum safe episode return of 70 faster than comparison methods. Safety depends on each experiment’s costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.