Separating physical prediction and visit-history error in Thinker’s imagined novelty
Thinker’s supplied exploration extension assigns imagined novelty by matching learned observation fingerprints against shared visit memory. We separate physical
Learning to explore with Thinker
AI · Reinforcement learning Thinker · NeurIPS 2023
Why Thinker stalls at lethal frontiers: novelty resolution, imagined novelty and termination, and the sign of death
Thinker's supplied exploration extension leaves the first room of symbolic Montezuma's Revenge but has never completed level 1. We ask why its exploration signa
Accomplishment Phase Conditioning, Prefix-Stratified Self-Imitation, and Hazard Planning in Thinker
Every policy reported to date on the symbolic Montezuma's Revenge Thinker benchmark (repeat=4, no sticky actions), including all twenty-three of our 6,000 s (~2
Learning to use retained action evidence in Thinker
Retaining an observation and predicting actions toward it does not establish that a learned planner can use the evidence to explore. We study this interface in