Reproduce “Q($γ_H$): Entropy-Guided Horizon Mixing in Multi-Discount Q-Learning” 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
Arwa Albelaihi, Abdulrahman Altahhan
Q(γH) adapts action selection across planning horizons by training value heads with fixed discount factors. At each state, it weights the heads using entropy-derived confidence in action values, while leaving each head’s Bellman update unchanged. It beats a matched uniform-weight baseline on MountainCar and records Acrobot-v1 AUC of −34,660 versus DDQN’s −37,900, with gains strongest on delayed-credit tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.