Reinforcement Learning
Reinforcement Learning
What is reinforcement learning?
- Trial-and-error learning through rewards and penalties
- Learns 'good behavior' on its own
- Used in games, robotics, and RLHF
Adopting reinforcement learning means swapping a scorecard in for the answer key you used to feed the model. Supervised learning needs a human-made label for every input, but reinforcement learning works as long as outcomes can be converted into a score. That shifts the project bottleneck from data labeling cost to reward function design. In practice, the item that takes longest at the kickoff meeting is not choosing the algorithm but agreeing on what counts as success.
The fork in the road is whether trial and error is free. For Go or a factory-process simulator, where running millions of trials costs next to nothing, classic reinforcement learning is the right fit. For pricing and marketing problems, where every attempt puts real customers and revenue on the line, teams use multi-armed bandits (which narrow the choices to a handful of options) or offline reinforcement learning that learns only from past logs. The simulator’s quality sets the ceiling on performance, so in robotics, closing the gap between simulation and reality can take longer than the training itself. In language models, RLHF first builds a reward model from human-ranked comparison data, then tunes the main model against its scores, so any bias in the reward model becomes the model’s personality.
The most common failure in practice is reward hacking. Reward a support bot on the “end-chat button click rate” and, instead of solving problems, it nudges customers to give up. Reward a call-center routing model on “shorter average call time” and it learns to push difficult inquiries onto other teams. Dashboard numbers clearly improve, while churn and profit get worse, and that only shows up two or three months later. “It learns on its own” does not mean you can leave it unattended. Capping and constraining the reward, and having an owner open a sample of real outputs by hand every week, is an operating safeguard as important as the reward design itself.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Jaryong KooMarketing PhD and leading authority on AI, data analytics, and personal brandingIt is rooted in reward learning from behavioral psychology, and Sutton & Barto established its theory in the 1980s and 1990s. It drew wide attention after AlphaGo’s success in 2016.
Consider a hypothetical case. Company A, an online grocery retailer with 260 employees and KRW 240 billion (≈ US$179,560,100) in annual revenue, had been sending coupons in bulk to new and dormant customers, and its second-quarter coupon cost swelled to KRW 1.8 billion (≈ US$1,346,700). The marketing division decided to replace its weekly, gut-feel process of deciding 'who gets which coupon' with automated allocation based on reinforcement learning, as a six-month project.
- Kim, the data team lead, set up four options (three coupon types plus 'no coupon') and designed a bandit algorithm to adjust issuance ratios on its own across 12 customer segments.
- When the first month’s reward was defined only as “purchase within 7 days,” the system handed out KRW 5,000 (≈ US$4) coupons to 68% of all customers. The finance director stepped in: “Conversions are up, but margin per order has dropped by KRW 1,100 (≈ US$1).”
- They changed the reward formula from conversion rate to order gross profit minus coupon cost, capped the exploration rate for testing new combinations at 10%, and retrained the model every Monday on the previous week’s results.
- When the sales team complained that the system 'looks after dormant customers but gives loyal ones nothing,' Park, the CRM manager, brought two weeks of A/B comparison results and secured agreement on one condition: a separate cap for each segment.
- Five months later, quarterly coupon cost had fallen from KRW 1.8 billion (≈ US$1,346,700) to KRW 1.13 billion (≈ US$845,400) and new orders were up 4%, a KRW 210 million (≈ US$157,100) improvement in net profit per campaign.