Answers are grounded in this paper and cite their sources.
Learning from experience in the real world requires reinforcement-learning rules that remain effective when environments change and actions incur economic costs. Disco103, a state-of-the-art self-discovered RL rule from games, provides a strong starting point for this transfer problem. Motivated by the era of experience and the big world hypothesis, we investigate how the organization of target-domain experience shapes adaptation under finite learning capacity. We separate experience organization from source-rule initialization using structured volatility-by-momentum tasks, a random-initialized control, and a pooled-experience control across 56 matched seed triplets and 168 meta-training runs. At the ten-bar decision interval, structured experience increases test Sharpe by 0.232 and compound annual growth by 1.58 percentage points over pooling, while reducing annual turnover by 68%, from 16.78 to 5.32 turns; all three effects have Holm-adjusted p_H < 0.001. These results establish experience organization as a first-class design choice in financial rule adaptation.
No takes yet. Share an insight, caveat, or question.
Luo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: