Update policy-evaluation.md

#Loss Function and Gradient: inconsistent usage of train sample. The role of the weigth factors is not clear: if n is the sample size, there is no need of weight factors. If the weigth factors are the frequency of state-action-target triple, then we need the frequency table of the sample including the target value (triple not pair) and n is not the sample size but the number of different triples.

content_md/reinforcement/approximate-methods/policy-evaluation.md

@@ -51,7 +51,7 @@ Each of these three tabular methods gives rise to an approximate value function
 
 ## Loss Function and Gradient
 
-Given training samples $(s_1,a_1,y_1),\ldots,(s_n,a_n,y_n)$ with inputs $(s_l,a_l)$ and targets $y_l$ we want to find the weigths $w$ of $Q_w$ by gradient descent. The loss function for training is usual mean squared error, but with additional weight factors $\mu_k$:
+Given training samples $(s_1,a_1,y_1),\ldots,(s_n,a_n,y_n)$ with inputs $(s_l,a_l)$ and targets $y_l$ we want to find the weigths $w$ of $Q_w$ by gradient descent. The loss function for training is usual mean squared error, but with additional weight factors $\mu_l$:
 \begin{equation*}
 L(w):=\sum_{l=1}^n\mu_l\,\bigl(Q_w(s_l,a_l)-y_l\bigr)^2
 \end{equation*}