Richard S. Sutton
= Richard S. Sutton
{c}
{wiki}
<Richard S. Sutton> is a <computer scientist> whose work developed <temporal-difference learning>, an important method in <reinforcement learning>. His 1988 paper studied how a predictor can learn from the difference between successive predictions, and proved convergence and optimality in specified special cases.<Learning to predict by the methods of temporal differences by Richard S. Sutton>[Abstract and sections 1–3]{ref}
A central idea is to update an earlier estimate using a later estimate before the eventual outcome is known. For example, if an estimated chance rises from $0.4$ to $0.7$, a learning rate of $0.2$ moves the earlier estimate toward the later one by $0.2(0.7-0.4)=0.06$. This illustrates the temporal credit-assignment idea; its guarantees depend on the learning problem and update rule, rather than following from this numerical example alone.
\Image[https://commons.wikimedia.org/wiki/Special:Redirect/file/Richard_Sutton%2C_October_27%2C_2016.jpg?width=260]
{title=Portrait of <Richard S. Sutton>, 27 October 2016}
{description=Steve Jurvetson; extracted from Deep Thinkers on Deep Learning. https://creativecommons.org/licenses/by/2.0/[CC BY 2.0], via Wikimedia Commons.}
{source=https://commons.wikimedia.org/wiki/File:Richard_Sutton,_October_27,_2016.jpg}
{width=260}