Policy Gradient Methods

Research Depth 73 in the knowledge graph I know this Set as goal
Unlocks 14 downstream topics
reinforcement-learning policy-optimization on-policy

Core Idea

Policy gradient methods directly optimize the policy π(a|s) via gradient ascent on expected return. REINFORCE uses full episode returns; advantage actor-critic uses value baselines. Methods are on-policy but handle continuous actions naturally.

Explainer

Most reinforcement learning methods you have seen so far work by estimating value functions — figuring out how good each state or action is, then deriving a policy indirectly by picking the highest-value action. Policy gradient methods take a fundamentally different approach: they parameterize the policy directly as a function π_θ(a|s) and optimize its parameters θ to maximize expected return. Instead of asking "what is the value of this action?" and choosing the best one, they ask "how should I adjust the probability of each action to get more reward?"

This direct approach solves a problem that value-based methods struggle with: continuous action spaces. If your agent controls a robotic arm with joint torques that can take any real-valued number, you cannot enumerate all possible actions to find the one with the highest Q-value. But a parameterized policy can output a probability distribution over continuous actions — for instance, a Gaussian with a learned mean and variance — and gradient ascent smoothly adjusts these parameters. Your background in gradient descent and partial derivatives applies directly here, except you are ascending (maximizing) the expected return J(θ) rather than descending a loss.

The key theoretical result is the policy gradient theorem, which gives a tractable expression for ∇_θ J(θ). The simplest algorithm built on it is REINFORCE: run a full episode under the current policy, compute the return G_t from each time step, and update θ in the direction of ∇_θ log π_θ(a_t|s_t) · G_t. Intuitively, this increases the probability of actions that led to high returns and decreases the probability of actions that led to low returns. The log-probability gradient tells you which direction in parameter space makes the chosen action more likely; the return G_t scales how strongly you push. REINFORCE is simple and unbiased, but it suffers from high variance because G_t depends on everything that happens after time t.

The standard remedy is to subtract a baseline from the return — typically a learned value function V(s_t). The quantity A_t = G_t − V(s_t) is called the advantage: it measures how much better the actual return was compared to the expected return from that state. If an action achieves average performance, its advantage is near zero and the policy barely changes. Only actions that perform surprisingly well or surprisingly poorly produce large updates. This is the actor-critic architecture: the "actor" is the policy π_θ, and the "critic" is the value function V that provides the baseline. The critic reduces variance without introducing bias (since subtracting a state-dependent baseline does not change the expected gradient), making learning significantly more stable and sample-efficient than raw REINFORCE.

Practice Questions 5 questions

Prerequisite Chain

Counting to 10Counting to 20Understanding ZeroThe Number ZeroCounting to FiveOne-to-One CorrespondenceCombining Small Groups Within 5Addition Within 10Addition Within 20Two-Digit Addition Without RegroupingTwo-Digit Addition with RegroupingAddition Within 100Repeated Addition as MultiplicationMultiplication Facts Within 100Division as Equal SharingDivision as Grouping (Measurement Division)Division: Grouping (Repeated Subtraction) ModelDivision: Fair Sharing ModelDivision as Equal SharingDivision as GroupingBasic Division FactsDivision Facts Within 100Two-Digit by One-Digit DivisionDivision with RemaindersRemainders and Quotients in DivisionDivision Word ProblemsIntroduction to Long DivisionFactors and MultiplesPrime and Composite NumbersEquivalent FractionsRelating Fractions and DecimalsDecimal Place ValueReading and Writing DecimalsComparing and Ordering DecimalsAdding and Subtracting DecimalsMultiplying DecimalsDividing DecimalsDividing FractionsMixed Number ArithmeticOrder of OperationsInteger Order of OperationsVariable ExpressionsCombining Like TermsOne-Step EquationsTwo-Step EquationsSolving Multi-Step EquationsEquations with Variables on Both SidesLiteral EquationsSlope-Intercept FormPoint-Slope FormWriting Linear EquationsParallel and Perpendicular Line SlopesGraphing Linear EquationsPiecewise FunctionsOne-Sided LimitsContinuity DefinitionLimit Definition of the DerivativePower RuleConstant Multiple and Sum/Difference RulesProduct RuleChain RuleHigher-Order DerivativesConcavity and Inflection PointsSecond Derivative TestCurve SketchingOptimization ProblemsCritical Points of Multivariable FunctionsCritical Points and Classification of ExtremaSecond Partial Test for Local Extrema (Hessian)The Hessian Matrix and Second Derivative TestUnconstrained Optimization: Finding ExtremaOptimization in Multiple VariablesIntroduction to Reinforcement LearningPolicy Gradient Methods

Longest path: 74 steps · 537 total prerequisite topics

Prerequisites (5)

Leads To (2)