Reinforcement learning frames a business decision, what price to charge, which offer to show, how to allocate a marketing budget, as a sequence of actions where an agent gets a reward after each one and gradually learns which actions tend to work best. There is no labelled training set of "correct" decisions here; the only way to find out how good a choice was is to make it and observe the outcome.
The core tension: explore or exploit
Exploitation means always choosing the action the agent currently believes is best, based on everything it has tried so far. Exploration means occasionally trying other actions specifically to gather more information, even ones that look worse right now, in case the current belief is wrong or has not been tested enough. An agent that only exploits risks getting permanently stuck on a mediocre choice that happened to look good early on, purely by chance.
Multi-armed bandits
A simplified version of this problem, called a multi-armed bandit, has several possible actions (arms) each with a fixed but unknown expected reward. A common strategy, epsilon-greedy, exploits the current best-known action most of the time but explores a random action with some fixed probability ε, gradually converging on the true best option as more data comes in.
Why this fits dynamic pricing
A pricing agent trying several candidate prices does not know upfront which one maximizes revenue, and customer response can shift over time. Framing pricing as a bandit problem lets the system continuously adapt, rather than relying on a static price set once and left unchanged.
Try it yourself
The Reinforcement Learning for Business Lab lets you tune the exploration rate on a five-price bandit problem and watch how quickly, and how profitably, the agent converges on the best price.
🧪 Try it yourself: the Reinforcement Learning for Business Lab simulation lets you experiment with everything described above directly in your browser.