Glossary
Action Space
The set of all valid actions in a given environment.
- discrete action spaces: only a finite number of moves are available to the agent
- continuous action spaces: actions are real-valued vectors
Advantage Function
The advantage function $A^{\pi}(s,a)$ corresponding to a policy $\pi$ describes how much better it is to take a specific action $a$ in state $s$, over randomly selecting an action according to $\pi(\cdot|s)$, assuming you act according to $\pi$ forever after.
$$ A^{\pi}(s,a) = Q^{\pi}(s,a) - V^{\pi}(s). $$
Markov Decision Processes
- https://spinningup.openai.com/en/latest/spinningup/rl_intro.html#optional-formalism
- https://www.geeksforgeeks.org/markov-decision-process/
- https://en.wikipedia.org/wiki/Markov_decision_process
An MDP is a 5-tuple, $\langle S, A, R, P, \rho_0 \rangle$, where
- $S$ is the set of all valid states,
- $A$ is the set of all valid actions,
- $R : S \times A \times S \to \mathbb{R}$ is the reward function, with $r_t = R(s_t, a_t, s_{t+1})$,
- $P : S \times A \to \mathcal{P}(S)$ is the transition probability function, with $P(s’|s,a)$ being the probability of transitioning into state $s’$ if you start in state $s$ and take action $a$,
- and $\rho_0$ is the starting state distribution.
The name Markov Decision Process refers to the fact that the system obeys the Markov property: transitions only depend on the most recent state and action, and no prior history.
Observation
Partial description of a state, which may omit information.
- fully observed: the agent is able to observe the complete state of the environment
- partially observed: the agent can only see a partial observation
Policy
A rule used by an agent to decide what actions to take.
- deterministic policy: a policy that always chooses the same action for a given state, with no randomness:
$$ a_t = \mu(s_t) $$
- stochastic policy: a policy that chooses actions probabilistically, introducing randomness: (e.g. categorical policies for discrete action spaces and diagonal Gaussian policies for continuous action spaces)
$$ a_t \sim \pi( \cdot \mid s_t) $$
- parameterized policy: policies whose outputs are computable functions that depend on a set of parameters (often denoted by $\theta$ or $\phi$)
$$ a_t = \mu_\theta (s_t) $$ $$ a_t \sim \pi_\theta ( \cdot \mid s_t) $$
Reward and Return
The reward function $R$ depends on the current state of the world, the action just taken, and the next state of the world:
$$ r_t = R(s_t, a_t, s_{t+1}) $$
The return is the cumulative reward over a trajectory, $R(\tau)$.
- finite-horizon undiscounted return: the sum of rewards obtained in a fixed window of steps:
$$ R(\tau) = \sum_{t=0}^T r_t $$
- infinite-horizon discounted return: the sum of all rewards ever obtained by the agent, discounted by how far off in the future they’re obtained. For $\gamma \in (0,1)$:
$$ R(\tau) = \sum_{t=0}^{\infty} \gamma^t r_t $$
State
A complete description of the state of the world.
Trajectory
A trajectory $\tau$ is a sequence of states and actions in the world.
$$ \tau = (s_0,a_0,s_1,a_1,…) $$
Value Function
A value function predicts the expected return from a state when following a specific policy.
- On-Policy Value Function $V^{\pi}(s)$: Expected return when starting in state $s$ and following policy $\pi$:
$$ V^{\pi}(s) = \underset{\tau \sim \pi}{\Epsilon}\left[R(\tau) \mid s_0 = s\right] $$
- On-Policy Action-Value Function: Expected return when starting in state $s$, taking action $a$, then following policy $\pi$:
$$ Q^{\pi}(s,a) = \underset{\tau \sim \pi}{\Epsilon}\left[R(\tau) \mid s_0 = s, a_0 = a\right] $$
- Optimal Value Function: Expected return when starting in state s and following the optimal policy
$$ V^* (s) = \max_{\pi}\underset{\tau \sim \pi}{\Epsilon}\left[R(\tau) \mid s_0 = s\right] $$
- Optimal Action-Value Function: Expected return when starting in state s, taking action a, then following the optimal policy
$$ Q^*(s,a) = \max_{\pi}\underset{\tau \sim \pi}{\Epsilon}\left[R(\tau) \mid s_0 = s, a_0 = a\right] $$