Exercise R.4.6

Probability mass functions: show the ε-greedy policy is a pmf

Reinforcement Learning › Exercises › Background review › R.4 Probability mass functions
EasyWritten

Is the following function a pmf for a finite action space \(\mathcal{A}\) and an action-value function \(Q_t(a)\)? \[\begin{align*} \pi_t(a)=\begin{cases} (1-\varepsilon)+\frac{\varepsilon}{ \mid \mathcal{A} \mid } &\quad\text{if $a=argmax \text{ } Q_t(a) $}\\ \frac{\varepsilon}{ \mid \mathcal{A} \mid } &\quad\text{otherwise}\\ \end{cases} \end{align*}\]