The authors examined long evaluation rollouts from Decision Transformers trained on offline D4RL datasets.
Prediction error helps Decision Transformers reject unreliable rollout contexts
TGDT filters trajectory-history suffixes using calibrated next-state prediction error before a critic ranks their proposed actions.
Big Tech
Chainesh Gautam · Raghuram Bharadwaj Diddigi · Chandramouli Kamanchi · Pankaj Dayama · Sumanta Mukherjee · Kameshwaran Sampath
International Institute of Information Technology Bangalore · IBM Research Bangalore
Research Digest··2 min read
The authors identify persistent next-state prediction error as a signal that a Decision Transformer's rollout context has drifted beyond its offline training distribution.
Why this paper
From IBM Research Bangalore and International Institute of Information Technology Bangalore
In one line
Decision Transformer's long rollout failures stem from context drift detectable by next-state prediction error; TGDT filters contexts by this error to improve returns.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§