Reinforcement Learning (RL)

Reinforcement learning (RL) is the machine-learning paradigm where an **agent** learns by interacting with an **environment**: it observes a state, chooses an action according to a **policy**, receives a **reward** and a new state, and updates its policy to maximise cumulative reward over time. The formal framework is the **Markov Decision Process** (MDP), which captures sequential decision-making under uncertainty. Unlike supervised learning, there are no labelled examples — the learning signal is the reward, which may be sparse, delayed, or only available after long action sequences.

The field's technical history spans classical tabular methods (Q-learning, SARSA, TD-learning), deep RL (DQN on Atari, AlphaGo's Monte Carlo Tree Search + policy network, policy-gradient methods), model-based RL (world models, MuZero), and offline RL (learning from fixed datasets without further environment interaction). The most consequential industrial impact of RL has come from its role in **RLHF** and related preference-based fine-tuning for language models, and from its role in robotics, game AI, autonomous driving simulators, and large-scale recommendation systems.

For APAC mid-market teams, most business applications do not require building an RL system from scratch. The relevant applications tend to be: (1) **personalisation and recommendations** where exploration-vs-exploitation matters (contextual bandits are often the right tool, not full RL); (2) **operations optimisation** in structured environments (inventory, pricing, ad spend, where classical tools plus a learning layer can beat pure optimisation); (3) **preference fine-tuning** of language models (DPO is usually the right choice, not classical PPO-based RLHF); (4) **simulation-driven design** where a model can be trained in a simulator before deployment.

The non-obvious operational warning: **reward hacking** is the most common RL failure mode. The agent finds a way to maximise the reward signal that does not correspond to the intended behaviour — a cleaning robot that hides messes rather than cleaning them, a boat-racing agent that loops a bonus region instead of finishing the race, a language model that games its preference-model rating. Reward design is harder than it looks, and more expertise than you expect is spent detecting and closing reward-hacking exploits. For business RL, keep the reward signal as close as possible to the metric you actually care about, and monitor for degenerate strategies.