Reinforcement learning

Markov decision process

The agent-environment interaction in a Markov Decision Process.

• Agent: decision maker

• Action: agent's decision which impacts the environment ($A_t$)

• State: the observation of the environment ($S_t$)

• Reward: a special numerical values ($R_t$)

• Environment: comprising everything outside the agent

At each time step $t \in \{0,1,2,...\}$, the agent receives some representation of the environment's state, $S_t \in \boldsymbol{S}$, and on that basis selects an action, $A_t \in A(s)$, which depends on the states. The agent receives a numerical rewards, $R_{t+1} \in R$.

The history is the sequence of states, actions, rewards:

$$S_0, A_0, R_1, S_1, A_1, R_2, S_2, A_2, R_2,...$$

The conditional probability of $s,r$ at time $t$ is like:

$$p(s',r|s,a) =P(S_{t}=s', R_{t}=r|S_{t-1}=s,A_{t-1}=a)$$

State-transition probability

$$p(s'|s,a) =P(S_{t}=s'|S_{t-1}=s,A_{t-1}=a) =\sum_{r \in R}p(s',r|s,a)$$

Expected rewards for state-action pairs

$$r(s,a) =\mathbb{E}[R_t|S_{t-1}=s,A_{t-1}=a] =\sum_{r \in R}r\sum_{s' \in S}p(s',r|s,a)$$

Expected rewards for state-action-next-state triples

$$r(s,a,s') =\mathbb{E}[R_t|S_{t-1}=s,A_{t-1}=a,S_t=s'] =\sum_{r \in R}r \dfrac{p(s',r|s,a)}{p(s'|s,a)}$$

• Return: the total reward an agent can get. ($G_t$)

• Policy: a probability distribution of actions under some states. ($\pi(a|s)$)

• Value: the expected return of a state under a policy. ($v_\pi(s)$)

Return or discounted return

$$G_t = \sum_{k=0}^{\infty}\gamma^k R_{t+k+1}, \qquad 0\le\gamma\le 1$$

The state-value function for policy $\pi$

$$v_\pi(s)=\mathbb{E}_\pi \left[ G_t | S_{t}=s \right]$$

The action-value function for policy $\pi$

$$q_\pi(s,a)=\mathbb{E}_\pi \left[ G_t | S_{t}=s , A_{t}=a \right]$$

The Bellman equation for $v_\pi(s)$

$$v_\pi(s)=\sum_a \pi(a|s) \sum_{r,s'}p(r,s'|s,a) \left[ r + \gamma v_\pi(s') \right]$$

The Bellman equation for $q_\pi(s,a)$

$$q_\pi(s,a) = \sum_{r,s'} p(r,s'|s,a) \left[ r + \gamma \sum_{a'} \left[ \pi(a'|s')q_\pi(s',a') \right] \right]$$

Optimal state-value function

$$v_*(s) = \max_\pi v_\pi(s)$$

Prediction & Control

• Prediction: estimating the value function for a given policy $\pi$

• Control: finding the optimal policy