3 ms·
For those who are not familiar, the I-SPY 2 trial is quite remarkable. It is an adaptive trial design - in essence reinforcement learning with human experiment
by pfisherman 3y ago
For those who are not familiar, the I-SPY 2 trial is quite remarkable. It is an adaptive trial design - in essence reinforcement learning with human experiments.[1]
For all the criticism FDA gets, they should also get credit for leaning into new approaches to testing drugs like adaptive trials. Part of why the US remains one of the most innovative countries.
1. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7731787/ https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7731787/
- hervature 3y agoIt should be clarified that adaptive trial design was studied for several decades before being applied to games and called reinforcement learning by computer scientists. The multi-armed bandit papers all reference "Asymptotically efficient adaptive allocation rules" because they derived the optimal lower bound for regret [1]. [1] - https://doi.org/10.1016/0196-8858(85)90002-8 https://doi.org/10.1016/0196-8858(85)90002-8
- smolder 3y agoI genuinely appreciate the link, but not so much the campaigning for the FDA or USA more broadly, which, my stance aside, feels like a trap leading to a fruitless discussion.