Imagine teaching a puppy to sit. You don't give it a textbook on sitting. Instead, you say "sit," and when it parks its rear on the floor, you give it a treat. If it just stares at you, it gets nothing. Over time, the puppy connects the action of sitting with the reward of a treat. It learns through trial and error.

This is the core idea behind reinforcement learning (RL). It's a way for a machine to learn by doing, figuring out the best actions to take in a situation to get the most reward.

In RL, we formalize this learning process with two main characters: the agent and the environment.

• The agent is the learner and decision-maker. It's our puppy.
• The environment is the world the agent interacts with. It's the room, you (the trainer), and the laws of physics.

The relationship is a continuous feedback loop. The agent observes the environment, takes an action, and the environment responds with a new observation and a reward.

This cycle repeats over and over. At each step, the agent performs an action. The environment then tells the agent what happened next (the new state) and whether that was good or bad (the reward). The agent's goal is to learn a strategy, or policy, for choosing actions that will rack up the highest possible total reward over time.

Let's break down the components of that loop: states, actions, and rewards.

For a robot cleaning a room, the state could be its location, its battery level, and the locations of all the dirt. For a stock-trading bot, the state could be the current prices of a set of stocks.

The set of all possible actions is called the action space. This can be simple, like 'up' or 'down', or incredibly complex, like all the possible moves in the game of Go.

Rewards are often immediate. The cleaning robot might get a +1 reward for sucking up dirt and a -10 reward for bumping into a wall. The ultimate goal isn't just to get the next reward, but to maximize the total, cumulative reward over the long run. Sometimes a short-term sacrifice (a small negative reward) can lead to a much larger future reward.

To tie all these concepts together, we use a framework called a Markov Decision Process (MDP). It's a mathematical way to describe the RL problem without getting bogged down in complex details. An MDP provides the rules of the game.

An MDP assumes something called the Markov Property: the future depends only on the current state, not on the sequence of events that preceded it. In other words, the current state contains all the necessary information to decide the next action. This is a powerful simplification. For our chess game, it means your best next move depends only on the current board layout, not how the pieces got there.

The transition probability $P$ and reward function $R$ define the dynamics of the environment. In many problems, the agent doesn't know these dynamics and has to learn them through exploration.

By framing a problem as an MDP, we can clearly define the agent's goal: find a policy (a strategy for choosing actions in each state) that maximizes the cumulative reward. Understanding this structure is the first step toward solving complex reinforcement learning challenges.