Warning

This page is located in archive.
Go to the latest version of this course pages.
Go the latest version of this page.

Consider the following MDP. Assume that reward is in the form $r(e,y)$, i.e., $r: E \times Y \mapsto \mathbb{R}$. Set $\gamma = \frac{1}{2}$. Suppose that you have seen the following sequence of states, actions, and rewards: $$ e_1, \mathrm{switch}, e_2, \mathrm{stay}, +1, e_2, \mathrm{stay}, +1, e_2, \mathrm{switch}, e_1, \mathrm{stay}, e_1, \mathrm{switch}, e_1, \mathrm{switch}, e_1, \mathrm{stay}, e_1, \mathrm{switch}, e_2, \mathrm{stay}, +1, {\color{lightgray} e_{\color{lightgray} 2}} $$

- What is $\hat{U}^{\pi}(e_i)$ calculated by the Direct Utility Estimation algorithm?
- What is transition model $\hat{\mathcal{E}}$ estimated by the Adaptive Dynamic Programming algorithm?
- What are state values estimated by a Temporal Difference learning agent after two steps? Assume that $\alpha=0.1$ and all values are initialized to zero.

Decide whether the following statement is true or false: * If a policy $\pi$ is greedy with respect to its own value function $U^{\pi}$, then this policy is an optimal policy.*

Do the following exploration/exploitation schemes fulfill the 'infinite exploration' and 'greedy in limit' conditions? Which lead to convergence of $Q$-values in $Q$-learning and which lead to convergence of $Q$-values in SARSA. Does anything change if we are interested in the convergence of policy? $N_{e,y}$ denotes the number of times when action $y$ was taken in state $e$. $N_e$ is defined similarly.

- a random policy
- $$ \pi(e) = \begin{cases} y, & \mbox{if } N_{e,y} \leq 100, \\ \mathop{\mathrm{arg\, max}}_y Q(e,y), & \mbox{otherwise}.\end{cases} $$
- $\varepsilon$-greedy policy with $\varepsilon = \frac{1}{N_e^2}$
- $\varepsilon$-greedy policy with $\varepsilon = \frac{1,000}{999+N_e}$
- $\varepsilon$-greedy policy with $\varepsilon = \frac{1}{\sqrt{N_e}}$

Consider the following MDP with $\gamma=0.8$, $r(5) = 100$, $r(\cdot) = 0$. The initial matrix of $Q$-values is $$ Q(e,y) = \begin{bmatrix} - & - & - & - & 0 & - \\ - & - & - & 0 & - & 0 \\ - & - & - & 0 & - & - \\ - & 0 & 0 & - & 0 & - \\ 0 & - & - & 0 & - & 0 \\ - & 0 & - & - & 0 & 0 \end{bmatrix}. $$ Consider path $1-5-1-3$ and constant learning rate $\alpha = 0.1$. Show changes in $Q$ values after the agent-environment interaction.

Problems 1,2, and 4 were adapted from Richard Sutton's 609 course. You may find the related materials on http://www.incompleteideas.net/book/the-book-2nd.html.

courses/smu/tutorials/tutorial9.txt · Last modified: 2019/04/15 12:22 by rysavpe1