Using the Kelly objective directly can destroy your RL network; practitioners use CRRA instead

rosstaylor90 · x · 2026-10-27

An RL practitioner shares how objective choice matters in practice:

Directly applying the Kelly objective to RL breaks badly: if the agent experiences ruin, the advantage goes to negative infinity under log utility, and optimization destroys the neural network — the 'ultimate doomer algorithm' that hates ruin so much it wrecks itself.
The practical fix: use a CRRA formulation with risk aversion tuned more conservative than risk-neutral but less extreme than Kelly, so RL dislikes ruin without halting training to avoid it.

More from Fun