The authors introduce predictive sufficiency, a decomposition of the Bayes-optimal next-state prediction error into components attributable to: (i) information recoverable from the agent's history, (ii) residual requiring the true context, and (iii) the finite-model deficit.
Predictive sufficiency quantifies when context encoders aid model-based RL agents
Authors show that the benefit of supplying true context for next-step prediction depends on how well the agent's own visitation pattern already reveals the latent dynamics.
Big Tech
Oleg Smirnov · Sofiane Ennadir · John Pertoft · Bjartur Hjaltason · Sara Karimi
King AI Labs, Microsoft Gaming · KTH Royal Institute of Technology
Research Digest··3 min read
The authors formalize 'predictive sufficiency' to measure what access to the true latent context adds over history alone for next-step prediction in model-based reinforcement learning.
Why this paper
From King AI Labs, Microsoft Gaming and KTH Royal Institute of Technology
In one line
Predictive sufficiency measures when an agent's history already identifies the environment context, making context encoders unnecessary.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§