Pinned
Long-horizon values depend on shorter-horizon values.
But what if those shorter-horizon values are still wrong?
We introduce DCRL: a recursive divide-and-conquer approach that learns values from short to long, substantially mitigating value-error accumulation over long horizons
