RL Policy: Agent Decision-Making Function | Generated by AI

Home 2026.05

Question: What is a “policy” in reinforcement learning?

Answer:

A policy π is simply a function that maps states to actions:

π(a | s) → probability of taking action a in state s

That’s it. It’s the agent’s decision-making function — the thing being learned.

Two flavors

Stochastic policy — outputs a probability distribution over actions:

probs = softmax(neural_net(state))  # [0.1, 0.7, 0.2]
action = sample(probs)              # → action 1

Deterministic policy — outputs a single action directly:

action = neural_net(state)          # → 0.73 (continuous control)

In LLM terms

The LLM is the policy:

π_θ(token | context) = softmax(W · transformer(context))

When you do RLHF/GRPO, you’re literally optimizing θ so the policy generates higher-reward token sequences. The “policy update” = gradient step on the weights.

The full RL loop

state s
    ↓
policy π(a|s)       ← this is what you're learning
    ↓
action a
    ↓
environment
    ↓
next state s', reward r
    ↓
use (s, a, r, s') to improve π
    ↓
repeat

Concrete minimal example

import torch
import torch.nn as nn

class Policy(nn.Module):
    def __init__(self, state_dim, action_dim):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(state_dim, 64),
            nn.ReLU(),
            nn.Linear(64, action_dim)
        )
    
    def forward(self, state):
        logits = self.net(state)
        return torch.distributions.Categorical(logits=logits)

policy = Policy(state_dim=4, action_dim=2)  # e.g. CartPole

state = torch.tensor([0.1, -0.2, 0.05, 0.3])
dist = policy(state)
action = dist.sample()           # stochastic action
log_prob = dist.log_prob(action) # needed for policy gradient

log_prob is the key quantity — the REINFORCE gradient is:

∇θ J = E[ ∇θ log π_θ(a|s) · R ]

You nudge the policy to make high-reward actions more probable.

One-line summary

Policy = the brain. State goes in, action comes out. RL = finding the best brain via trial and error.


Back Donate