Pearl
Start

Chapter 00 · Prelude

Why do we need MDPs?

Abstract. Before Markov Decision Processes, people already tried to make machines decide. Bandits, open-loop control, supervised “behavior cloning,” classical planners, and reactive lookup tables all worked in narrow settings — and broke as soon as the world became sequential, uncertain, and long-horizon. This chapter is the motivation for Chapter 01: what we needed that those tools did not give us.

1. Why do we need MDPs?

Reinforcement learning is not “train a network on labels.” It is the problem of an agent that acts over time, gets sparse feedback, and must live with the consequences of earlier choices.

That problem has three stubborn features:

  1. Sequential dependence. Today’s action changes tomorrow’s state.
  2. Delayed credit. A reward now may be caused by a choice many steps ago.
  3. Uncertainty. The same action in the same situation need not always produce the same next state.

An MDP is a mathematical contract that names those features cleanly: states, actions, transitions, rewards, and a discount factor. Without that contract, algorithms have nowhere stable to stand — you cannot even write down what “optimal” means.

Agent interacts with environment, receiving state and reward
Agent–environment loop. Wikimedia Commons, Megajuice, CC0.
The goal, in one line

Find a policy \(\pi\) that maximizes expected long-run return \(\mathbb{E}_\pi[G_0]\). Everything after Chapter 00 is machinery for that sentence.

2. What came before MDPs

People did not wait for Bellman to invent decision-making. Several older frameworks already tried to answer “what should I do?”

Multi-armed bandits

Pull one arm; get a reward; repeat. Exploration vs exploitation is real — but there is no lasting state. One action does not rearrange the world for later.

no dynamics · only immediate reward

Supervised learning

Map \(x\mapsto y\) from labeled data. Brilliant at perception. Silent about “if I take this action now, what happens three steps later?”

labels · i.i.d. assumption

Open-loop control

Compute a fixed sequence \(a_0,a_1,\ldots\) in advance and play it. Works if the world is nearly deterministic and nothing goes wrong. Brittle the moment noise appears.

plan once · never replan

Classical planning (deterministic)

Search a graph of known, deterministic transitions to a goal. Excellent when \(P\) is known and sharp. Collapses when outcomes are probabilistic or the model is wrong.

known graph · no stochastic P

Reactive / scripted policies

Hand-written rules: “if near wall, turn left.” Fast and interpretable, but they do not optimize long-horizon return and do not generalize when the rule book is incomplete.

if–then · no optimality

Behavior cloning

Imitate expert actions with supervised learning. Fails under distribution shift: the learner drifts off the expert’s state distribution and never recovers (compounding error).

imitate · no recovery

None of these are “wrong”

Bandits, SL, planning, and scripts are still useful tools. They failed as a complete theory of sequential decision-making under uncertainty — which is exactly the setting RL targets.

3. Why those ideas failed (for this problem)

Bandits ignore the future world

In a bandit, pulling arm 3 does not change which arms exist tomorrow. In a maze, moving right does change where you are. Credit assignment across time is not optional — it is the whole problem.

Supervised learning has no consequence model

A classifier answers “what label is this?” An agent must answer “what happens if I act?” Without transitions and rewards, there is no notion of a better or worse future — only better or worse predictions of someone else’s labels.

Open-loop plans shatter under noise

Suppose the plan says “go right, right, up.” If the first right slips into a pit, the remaining plan is nonsense. Closed-loop decisions — actions that depend on the current state — are mandatory. That is a policy \(\pi(a\mid s)\), not a fixed tape of actions.

Deterministic planners refuse probability

Real robots slip. Dice roll. Opponents move. If your math assumes \(P(s'\mid s,a)\in\{0,1\}\), then either you pretend the world is sharp or you invent ad-hoc patches. MDPs put the distribution in the model from day one.

Imitation compounds error

Behavior cloning trains \(\hat\pi(a\mid s)\) on expert states. Once the learner takes a slightly wrong action, it visits states the expert never showed — and the trained policy has no preference signal there. Without rewards or a recovery objective, small mistakes become catastrophic trajectories.

Failure mode in one sentence

Pre-MDP tools either ignored dynamics, ignored uncertainty, ignored delayed reward, or ignored closed-loop feedback — and sequential agents need all four.

4. The missing pieces

To state the sequential decision problem properly we need names for:

PieceWhy it matters
State \(s\)What the agent knows when it decides
Action \(a\)What it can choose
Transition \(P(s'\mid s,a)\)How the world reacts (possibly randomly)
Reward \(R\)Scalar feedback for preference
Discount \(\gamma\)How much future matters vs now
Policy \(\pi(a\mid s)\)Closed-loop strategy, not open-loop tape
Return \(G_t\)Long-horizon score to maximize in expectation

Pack those into one object and you have an MDP. That is not fashion — it is the smallest language in which “optimal sequential behavior under uncertainty” is even a well-posed question.

5. What the MDP buys us

Once \(\mathcal{M}=(S,A,P,R,\gamma)\) is on the table:

  • We can define value functions \(V^{\pi}\) and \(Q^{\pi}\) (Chapter 02).
  • We get Bellman equations — local consistency of long-horizon return.
  • We can run Dynamic Programming when \(P\) is known (Chapter 03).
  • We can learn from samples when \(P\) is unknown (TD, Q-learning, later chapters).

Every later algorithm is still negotiating with this same object. If the MDP assumption is wrong (non-Markov sensors, missing state), the mathematics quietly lies — which is why state design matters as much as the optimizer.

Takeaway

We need MDPs because earlier frameworks could not jointly express dynamics + uncertainty + delayed reward + closed-loop policies. The MDP is the language. RL is what we do when we speak that language — with a model, or from experience.

6. What follows

Chapter 01 writes the MDP down carefully: the tuple, the Markov property, returns, and policies. Then values and Bellman equations turn the definition into computation.

See also

Next chapter

Markov Decision Processes

States, actions, \(P\), \(R\), \(\gamma\) — the contract itself.

Continue →