Rylan Schaeffer

Kernel Papers

Rescorla-Wagner Learning Rule

The Rescorla-Wagner Learning Rule (1972) was a seminal model of associative learning that preceded reinforcement learning. Associative learning is the problem of learning how different stimuli are associated with rewards or punishments \(r_n\), where \(n\) indexes the trial number. The model considers the agent receiving a one-hot encoded stimulus vector \(s_n\), where each element indicates the presence or absence of a stimulus and \(n\) is the trial number, and the agent then uses a linear readout \(w_n\) of the stimuli to predict the expected reward or punishment \(v_n\):

\[v_n = w_n^T s_n\]

Over the course of the $N$ trials, the linear readout $w_n$ is updated using the prediction error, $r_n - v_n$ (occasionally denoted by $\delta_n$):

\[w_{n+1} \leftarrow w_n + \eta (r_n - v_n) s_n\]

This learning rule is equivalent to online gradient descent under a mean-squared error loss between the actual reward and the expected reward:

\[\begin{align*} L(w) &= \langle (r - v)^2 \rangle_{s} \\ \nabla_w L(w) &= \langle 0 - 2 r s + 2 s s^T w \rangle_{s}\\ &= 2 \langle (r - w^T s) s \rangle_{s}\\ &\propto \langle (r - v) s \rangle_{s} \end{align*}\]