Reinforcement Learning

Reinforcement Learning (RL) #

RL is learning by trial and error.

Reinforcement Learning (RL) is a type of machine learning where an autonomous agent learns to make decisions by interacting with an environment.

Instead of being told the correct answer, the agent:

  • takes actions
  • observes outcomes
  • receives rewards or penalties
  • gradually learns a strategy that maximises long-term reward

Reinforcement Learning teaches an agent how to act, not what to predict.

flowchart LR
    A[Agent] -->|Action| B[Environment]
    B -->|State + Reward| A

1. Agent #

  • The learner and decision-maker
  • Chooses actions based on its current policy

2. Environment #

  • Everything the agent interacts with
  • Responds to actions by changing state and giving rewards

3. State #

  • A snapshot of the environment at a given time

4. Action #

  • A choice made by the agent

5. Reward #

  • Feedback signal
  • Positive reward → good outcome
  • Negative reward → bad outcome

6. Policy #

  • The agent’s strategy
  • A mapping from states to actions
  • Goal: maximise cumulative reward over time

What Is an Autonomous Agent? #

An autonomous agent is a system that:

  • observes its environment
  • makes decisions independently
  • acts without direct human instruction

Examples #

  • Robots
  • Self-driving cars
  • Game-playing agents (Chess, Go, Atari)
  • Recommendation systems

How Reinforcement Learning Differs from Other ML Types #

Reinforcement Learning vs Supervised Learning #

Supervised Learning

  • Uses labelled data
  • Learns from examples of correct answers
  • Goal: minimise prediction error

Reinforcement Learning

  • No labelled “correct” actions
  • Learns from rewards and penalties
  • Goal: maximise long-term reward

RL is not told what the right action is — it must discover it.

Reinforcement Learning vs Unsupervised Learning #

Unsupervised Learning

  • Finds hidden patterns in data
  • No explicit objective tied to actions
  • Focuses on structure (clusters, embeddings)

Reinforcement Learning

  • Learns by interaction
  • Uses a reward function
  • Focuses on decision-making

RL learns what to do, not just what exists.

A Key Conceptual Difference #

  • Supervised & Unsupervised Learning:

    • Assume data points are independent
    • Learn a static model from a dataset

  • Reinforcement Learning:

    • Data arrives as sequences
    • State → Action → Reward → Next State
    • Decisions affect future data

Reinforcement learning models sequential decision-making, which closely mirrors how humans and animals learn.

Reinforcement Learning from Human Feedback (RLHF) #

RLHF is a modern extension of reinforcement learning where humans help define what “good” looks like.

How RLHF Works #

  1. Humans evaluate model outputs (e.g., ranking answers)
  2. A reward model is trained from this feedback
  3. Reinforcement learning optimises the agent using this reward model

Why RLHF Matters #

Some concepts are hard to define mathematically:

  • humour
  • helpfulness
  • politeness
  • safety

Humans can easily judge these qualities, even if they can’t formalise them.

Example #

It’s hard to write a formula for “funny”, but easy for humans to rate jokes.
That feedback becomes a reward signal to improve the model.

RLHF is widely used in:

  • Large Language Models
  • Chatbots
  • AI assistants

Where Reinforcement Learning Is Used #

  • Robotics
  • Autonomous vehicles
  • Game-playing AI
  • Recommendation systems
  • Resource allocation
  • Training LLM behaviour (via RLHF)

Summary #

  • Reinforcement Learning trains agents, not predictors
  • Learning happens through trial and error
  • Rewards guide behaviour instead of labels
  • Decisions are sequential and interdependent
  • RLHF brings humans into the reward loop

Reinforcement Learning teaches machines how to behave by rewarding good decisions and punishing bad ones.

References #

Home | Machine Learning

Previous Semi-Supervised Learning Machine learning Workflow Next