Reinforcement learning (RL) studies sequential decision-making in which an agent observes information, selects actions, and receives a reward signal. Its objective is to learn a policy that maximizes expected cumulative reward under the training formulation—not to discover universally optimal or safe behavior.

RL has produced important demonstrations in games, robotics, recommendation research, operations, and scientific optimization. Evidence does not transfer automatically from a simulator to the real world. Deployment must address reward misspecification, distribution shift, unsafe exploration, partial observability, constraints, human behavior, and costly errors.

1. Games

Games offer explicit actions, rules, observations, and outcomes. AlphaGo combined policy and value networks, Monte Carlo tree search, supervised learning from expert games, and reinforcement learning from self-play. AlphaGo Zero and AlphaZero later reduced reliance on human game data. OpenAI Five and AlphaStar were also scoped self-play demonstrations.

These results establish capability in specified environments; they do not establish general reasoning, freedom from bias, or safe transfer to unrelated domains.

2. Automated driving research

RL can be investigated for motion planning, control, or decision policies in simulation. Simulator assumptions about sensors, actors, traffic, weather, and physics can differ from public roads. Do not train through unconstrained exploration on public roads. Validation needs a defined operational design domain, scenario and closed-course testing, independent safety analysis, staged exposure, monitoring, incident investigation, and an appropriate fallback.

3. Trading and execution research

Order execution and portfolio decisions can be formulated sequentially, but historical backtests are not live-market evidence. They are vulnerable to leakage, survivorship bias, overfitting, unrealistic fills, transaction cost, market impact, capacity, regime change, and participant feedback. Compare with simple baselines on time-ordered holdouts and enforce exposure, loss, liquidity, compliance, and kill-switch controls.

4. Robotics

RL has been studied for grasping, locomotion, and manipulation. Simulation and domain randomization expose a policy to variation, but sim-to-real transfer remains empirical. Validate sensor and actuator limits, collision and force constraints, human proximity, hardware wear, latency, emergency stop, fallback, and recovery. OpenAI’s Dactyl was a scoped research demonstration, not proof that arbitrary tasks transfer directly.

5. Recommendations and resource allocation

Contextual bandits and RL can optimize sequences of recommendations or allocations, but clicks and watch time are incomplete proxies for user welfare. Evaluation should include delayed effects, novelty, diversity, concentration, exposure fairness, user control, manipulation risk, and feedback loops. Offline policy evaluation depends on logged propensities, overlap, assumptions, and uncertainty; it is not a substitute for safe experimentation.

6. Operations and energy systems

RL has been researched for scheduling, inventory, cooling, traffic signals, and grid control. A paper or pilot does not establish broad production readiness. Compare against operations-research and rule-based baselines under forecast error, rare events, maintenance, cyber incidents, capacity constraints, and operator intervention. Safety-critical actions require hard constraints outside a learned reward.

7. Science and healthcare

RL can help search molecular designs, experimental sequences, or treatment-policy hypotheses. AlphaFold is not an RL application merely because it uses machine learning. Clinical observational data create confounding, censoring, missingness, and policy-shift problems; a retrospective policy estimate is not evidence that a treatment improves outcomes. Prospective validation, clinical oversight, ethics review, patient protections, and regulatory obligations still apply.

8. Language-model post-training

Reinforcement learning from human or AI feedback can adjust language-model behavior using preference data and a reward or preference model. It does not guarantee truthfulness, harmlessness, or alignment. Reward-model errors, evaluator disagreement, distribution shift, and specification gaming require independent evaluation. See LLM evaluation metrics.

A deployment checklist

  1. Define the decision process, affected people, baseline, objective, constraints, and prohibited actions.
  2. Separate published demonstrations from verified production evidence.
  3. Test reward misspecification, out-of-distribution states, uncertainty, subgroup impacts, and adversarial conditions.
  4. Use offline, simulated, shadow, or tightly constrained trials before live actions.
  5. Provide monitoring, human authority, safe fallback, rollback, and incident response.
  6. Version the policy, environment, data, reward, code, evaluation, and approvals through AI model management.

Explanations can support review but do not prove safety or causality; see explainable AI examples. The relevant question is not whether RL is impressive, but whether a versioned system improves a defined decision under realistic constraints with acceptable residual risk.

Originally published June 30, 2025; technically reviewed and substantially updated September 4, 2026.