The Power of RL

The Power of Reinforcement Learning

Continuous Learning for Better Predictions

TL;DR
Reinforcement Learning (RL) is a powerful way to build models that learn by doing. Rather than merely fitting historical data, RL optimizes decisions through rewards and feedback loops—from real production environments and simulations. The result: models that continue to improve continue to improve as the world changes. Think of applications ranging from AlphaGo-level decision-making to revenue and profit optimization, inventory and pricing strategies, and even stock market signaling (with the right governance).

  • Agent: the model that makes decisions.

  • Environment: the world in which the model operates (marketplace, online store, supply chain, stock exchange).

  • Reward: a number indicating how good an action was (e.g., higher margin, lower inventory costs).

  • Policy: a strategy that chooses an action given a state.

Acronyms explained:

  • RL = Reinforcement learning

  • MDP = Markov decision process (mathematical framework for RL)

  • MLOps = Machine learning operations (operational side: data, models, deployment, monitoring)

Why RL Is Relevant Now

  1. Continuous learning: RL adapts its policy when demand, prices, or behavior change.

  2. Decision-oriented: Not only predicting, but actually optimizing from the outcome.

  3. Simulation-friendly: You can safely run “what-if” scenarios before going live.

  4. Feedback first: Use real KPIs (margin, conversion, inventory turnover) as the direct reward.

Important: AlphaFold is a deep-learning breakthrough in protein folding; it the quintessential RL example is AlphaGo/AlphaZero (decision-making with rewards). The point remains: learning through feedback delivers superior policies in dynamic environments.
AlphaFold uses a combination of Generative AI to predict a GEN combination instead of predicting word combinations (tokens). It uses Reinforcement Learning to predict the most likely form of a given protein structure.

Business Use Cases (with a Direct KPI Link)

1) Optimize revenue and profit (pricing + promotions)

  • Objective: maximum gross margin with stable conversion.

  • State: time, inventory, competitor pricing, traffic, history.

  • Action: choose a price step or promotion type.

  • Reward: margin – (promotion costs + return risk).

  • Bonus: RL prevents overfitting to historical price elasticity because it explores.

2) Inventory & supply chain (multi-echelon)

  • Objective: service level ↑, inventory costs ↓.

  • Action: adjust reorder points and order quantities.

  • Reward: revenue – inventory and backorder costs.

3) Allocate the marketing budget (multi-channel attribution)

  • Objective: maximize ROAS/CLV (Return on ad spend / Customer lifetime value).

  • Action: allocate budgets across channels & creatives.

  • Reward: attributed margin in both the short and long term.

4) Finance & stock signaling

  • Objective: risk-adjusted maximize returns.

  • State: price features, volatility, calendar/macro events, news/sentiment features.

  • Action: position adjustment (increase/decrease/neutralize) or “no trade”.

  • Reward: PnL (Profit and Loss) – transaction costs – risk penalty.

  • Note: no investment advice; ensure strict risk limits, slippage models and compliance.

The Mantra LOOP:

Analyze → Train → Simulate → Operate → Evaluate → Retrain

This is how we ensure ongoing learning at Fortis AI:

  1. Analysis (Analyze)
    Data audit, KPI definition, reward design, offline validation.

  2. Train
    Policy optimization (e.g., PPO/DDDQN). Determine hyperparameters and constraints.

  3. Simulate
    Digital twin or market simulator for what-if and A/B scenarios.

  4. Operate
    Controlled rollout (canary/gradual). Feature store + real-time inference.

  5. Evaluate
    Live KPIs, drift detection, fairness/guardrails, risk measurement.

  6. Retrain
    Periodic or event-driven retraining with fresh data and outcome feedback.

Minimal pseudocode for the loop

while True:
data = collect_fresh_data() # realtime + batch
policy = train_or_update_policy(data) # RL update (bijv. PPO)
results_sim = simulate(policy) # sandbox/AB-test in simulator
if passes_guardrails(results_sim):
deploy(policy, mode="canary") # klein percentage live
kpis = monitor(realtime=True) # marge, conversie, risk, drift
if drift_detected(kpis) or schedule_due():
continue # retrain-trigger

Why RL Instead of “Just Predicting”?

Classical supervised models predict an outcome (e.g., revenue or demand). But the best prediction does not automatically lead to the best action. RL optimizes directly over the decision space with the actual KPI as the reward—and learns from the consequences.

In short:

  • Supervised: “What is the probability that X will happen?”

  • RL: “Which action maximizes my objective now and in the long term?”

Success Factors (and Pitfalls)

Design the reward carefully

  • Combine short-term KPIs (daily margin) with long-term value (CLV, inventory health).

  • Add penalties for risk, compliance, and customer impact.

Limit exploration risk

  • Start in simulation; go live with canary releases caps (e.g., maximum price movement per day).

  • Build guardrails: stop-losses, budget limits, approval flows.

Prevent data drift & leakage

  • Use a feature store with version control.

  • Monitor drift (statistics change) and retrain automatically.

Set up MLOps & governance

  • CI/CD for models, reproducible pipelines, explainability and audit trails.

  • Align with DORA/IT governance and privacy frameworks.

How Do You Get Started Pragmatically?

  1. Choose a KPI-focused, well-defined case (e.g., dynamic pricing or budget allocation).

  2. Build a simple simulator with the key dynamics and constraints.

  3. Start with a safe policy (rule-based) as a baseline; then test the RL policy side by side.

  4. Measure live, on a small scale (canary), and scale up after demonstrated uplift.

  5. Automate retraining (schedule + event triggers) and drift alerts.

What Fortis AI Delivers

For Fortis AI we combine strategy, data engineering, and MLOps with agent-based RL:

  • Discovery & KPI Design: rewards, constraints, risk limits.

  • Data & Simulation: feature stores, digital twins, A/B framework.

  • RL Policies: from baseline → PPO/DDQN → context-aware policies.

  • Production-Ready: CI/CD, monitoring, drift, retraining, and governance.

  • Business Impact: focus on margin, service level, ROAS/CLV, or risk-adjusted P&L.

Would you like to know which continuous learning loop that delivers the most value for your organization?
👉 Schedule an exploratory call via fortis ai.nl – we would be happy to show you a demo of how you can apply Reinforcement Learning in practice.

Gerard

Gerard works as an AI consultant and manager. With extensive experience at large organizations, he can unravel a problem and work toward a solution remarkably quickly. Combined with his economics background, this enables him to make commercially sound decisions.