Lecture 7.2 Reinforcement Learning Terminology

Policy

A policy maps states to actions, either deterministically or stochastically:

Deterministic: $\pi :\mathcal{S}\to \mathcal{A}$ Stochastic: $\pi :\mathcal{S}\to P(\mathcal{A})$

State Transitions

The dynamics/transition function specifies the probability of next states:

$P(s^{\prime }|s,a):\mathcal{S}\times \mathcal{A}\times \mathcal{S}\to [0,1]$

State Utility Function

The utility function mapping from state space to expected rewards (also denoted as V(s) for “value” in many texts, but following Russel & Norvig):

$U:\mathcal{S}\to \mathbb{R}$

Quality Function (Q-Function)

The quality function mapping from state-action pairs to expected total rewards:

$Q:\mathcal{S}\times \mathcal{A}\to \mathbb{R}$

Core Definitions

Return (Reward-to-go)

The discounted sum of future rewards from time step t:

$G_t={\sum }_{k=0}^{\infty }{\gamma }^k{r}_{t+k}$

Expected Utility Under a Policy

The expected return when following policy $\pi$:

$U^{\pi }(s)={\mathbb{E}}_{\pi }[{\sum }_{k=0}^{\infty }{\gamma }^k{r}_{t+k}|s_t=s]$

Quality Function Under a Policy

Expected return starting with action a and following policy $\pi$:

$Q^{\pi }(s,a)={\mathbb{E}}_{\pi }[{\sum }_{k=0}^{\infty }{\gamma }^k{r}_{t+k}|s_t=s,a_t=a]$

Advantage Function

The relative advantage of action a compared to the average action in state s:

$A^{\pi }(s,a)=Q^{\pi }(s,a)-U^{\pi }(s)$

Bellman Equations

Bellman Equation for Utility

$U^{\pi }(s)={\sum }_a\pi (a|s){\sum }_{s^{\prime }}P(s^{\prime }|s,a)[r(s,a,s^{\prime })+\gamma U^{\pi }(s^{\prime })]$

Bellman Equation for Q-Values

$Q^{\pi }(s,a)={\sum }_{s^{\prime }}P(s^{\prime }|s,a)[r(s,a,s^{\prime })+\gamma {\sum }_{a^{\prime }}\pi (a^{\prime }|s^{\prime })Q^{\pi }(s^{\prime },a^{\prime })]$

Bellman Optimality Equations

$U^{\ast }(s)={\max }_a{\sum }_{s^{\prime }}P(s^{\prime }|s,a)[r(s,a,s^{\prime })+\gamma U^{\ast }(s^{\prime })]$ $Q^{\ast }(s,a)={\sum }_{s^{\prime }}P(s^{\prime }|s,a)[r(s,a,s^{\prime })+\gamma {\max }_{a^{\prime }}{Q}^{\ast }(s^{\prime },a^{\prime })]$

Key Relationships

Policy Improvement

The greedy policy with respect to Q:

${\pi }^{\prime }(s)=\arg {\max }_a{Q}^{\pi }(s,a)$

Utility-Quality Relationship

$U^{\pi }(s)={\mathbb{E}}_{a\sim \pi (s)}[Q^{\pi }(s,a)]$ $U^{\ast }(s)={\max }_a{Q}^{\ast }(s,a)$

Learning Paradigms

Model-Free vs Model-Based

Model-Based

Requires transition function P(s’|s,a) and reward function r(s,a,s’)

Model-Free

Works directly with samples (s, a, r, s’) without requiring transition/reward models Uses sample estimates: $\hat{Q}(s,a)\approx Q^{\ast }(s,a)$

Active vs. Passive Learning

Passive Learning

• Policy $\pi$ is fixed
• Goal: Learn $U^{\pi }$ or $Q^{\pi }$ for the given policy
• Example: Policy evaluation

Active Learning

• Agent can modify policy $\pi$
• Goal: Find optimal policy ${\pi }^{\ast }$
• Examples: Q-learning, policy gradient methods

On-Policy vs. Off-Policy Learning

On-Policy Learning

• Learn about policy $\pi$ from experience generated by $\pi$
• Policy being evaluated = Policy generating experience
• Example: SARSA update: $Q(s_t,a_t)\leftarrow Q(s_t,a_t)+\alpha [r_t+\gamma Q(s_{t+1},a_{t+1})-Q(s_t,a_t)]$ where $a_{t+1}$ is chosen according to current policy $\pi$

Off-Policy Learning

• Learn about target policy $\pi$ from experience generated by behavior policy $\mu$
• Enables learning optimal policy while following exploratory policy
• Example: Q-learning update: $Q(s_t,a_t)\leftarrow Q(s_t,a_t)+\alpha [r_t+\gamma {\max }_{a^{\prime }}Q(s_{t+1},a^{\prime })-Q(s_t,a_t)]$
• Important concept: Importance Sampling Ratio: ${\rho }_t=\frac{\pi (a_t|s_t)}{\mu (a_t|s_t)}$ Used to correct for difference between target and behavior policies