Jethro's Braindump

Markov Decision Process

A sequential decision problem for a fully observable, stochastic environment with a Markovian transition model and additive rewards is called a Markov decision process. It consists of a set of states, with initial state \(s_0\), a set \(ACTIONS(s)\) of actions in each state, a transition model \(P(s'|s, a)\), and a reward function \(R(s)\).

A policy, denoted \(\pi\), specifies what the agent should do in any state \(s\). This action is denoted by \(\pi(s)\). The optimal policy \(\pi^*\) yields the highest expected utility.

The careful balancing of risk and reward is a characteristic of MDPs that does not arise in deterministic search problems.

Optimality in Markov Decision Processes

Finite Horizon

\begin{equation} E\left( \sum_{t=0}^{h} r_t \right) \end{equation}

Infinite Horizon

\begin{equation} E\left( \sum_{t=0}^{\infty} \gamma^t r_t \right) \end{equation}

Average-reward

\begin{equation} \lim_{h \rightarrow \infty} E\left( \sum_{t=0}^{h} \frac{1}{h} r_t \right) \end{equation}

Learning Performance (Kaelbling, Littman, and Moore, n.d.)

  1. Asymptotic convergence:

\begin{equation} \pi_n \rightarrow \pi^* \text { as } n \rightarrow \infty \end{equation}

  1. PAC:

\begin{equation} P(N_{errors} > F(\cdot, \epsilon, \delta)) \le \delta \end{equation}

Does not give any guarantee about the policy while it is learning

  1. Regret (e.g. bound \(B\) on total regret):

\begin{equation} \mathrm{max} \sum_{t=0}^{T} r_{tj} - r_t < B \end{equation}

No notion of “many small mistakes”, or “few major mistakes”.

  1. Uniform-PAC

unifies notion of PAC and regret (Dann, Lattimore, and Brunskill, n.d.)

Bibliography

Dann, Christoph, Tor Lattimore, and Emma Brunskill. n.d. “Unifying PAC and Regret: Uniform PAC Bounds for Episodic Reinforcement Learning.” In Advances in Neural Information Processing Systems, 5713–23.

Kaelbling, Leslie Pack, Michael L Littman, and Andrew W Moore. n.d. “Reinforcement Learning: A Survey” 4:237–85.

Links to this note