Kevin Murphy unifies goal-based hierarchical RL with ACGVF and his 25-year-old HHMM work

sirbayes · x · 2026-09-15

Kevin Murphy released a mini-paper, "A note on goal-based hierarchical RL", unifying Tasse et al.'s agent-centric general value function (ACGVF) construction — which lets the agent choose which goal to pursue and when a goal is finished — with his own 25-year-old work on hierarchical hidden Markov models (HHMMs). ACGVF subsumes nearly all prior RL, control and planning formalisms but assumes fully observed environments; Murphy's earlier belief-state agent design assumed externally provided goals. The note extends both via the HHMM formalism, though no experiments yet.

Original post →

More from Research

Research channel →