Reinforcement Training

1) Purpose and place of RL in iGaming

• Personalization of the game catalog (Slate-RL): selection of a set of offers for the screen/push.
• Bonus/promo optimization: size/type/timing taking into account the risk of abuse.
• Reactions in RG/Retention: when and how to intervene (soft notifications/pause/escalation).
• Operations: dynamic limit management, prioritization of support queues.
• Traffic and procurement: bidding in auctions, budget-pacing.

Why not only supervised: the target variable is a long-term reward (LTV, wellbeing, risk reduction), which must be optimally accumulated, and not just predicted.

2) Basic wording

Status (s_t): player profile, session context, market restrictions.
Action (a_t): offer, selection of games (slate), RG trigger, bidder bet.
Reward (r_t): mixed metric (income - RG/AML penalties - value).

3) Method families

3. 1 Bandits (stateless)

Multi-Armed Bandit: (\epsilon)-greedy, UCB, Thompson Sampling.
Contextual bandits: Consider player/session traits.
Slate/Ranking Bandits: Offer Set Selection; adjust positional effects.

3. 2 Full RL

Policy Gradient/Actor-Critic: REINFORCE, A2C/A3C, PPO - resistant to large spaces.
Q-Learning/Deep Q-Networks: Discrete Actions, Offline Learning with Experience Buffer.
Conservative/Offline RL: CQL, BCQ, IQL - learn from logs without online exploitation.

3. 3 Safe/Restricted RL

Constrained RL (CMDP): optimization under RG/AML/budget constraints.
Risk-Sensitive: CVaR-RL, penalty shaping, Lagrangian методы.

4) Award design (reward shaping)

• Revenue: Contribution to incremental Net Revenue/LTV (not "raw rates").
• Responsible play: penalties for risk patterns, exceeding limits, tiring incentives.
• Compliance/AML: penalties for actions that increase the likelihood of unsafe activity.
• Experience quality: CTR/CVR/session length, but with cap/weight to avoid "overheating."

python reward = w_revdelta_net_revenue \
- w_rgrg_risk_score \
- w_abusebonus_abuse_prob \
- w_costoffer_cost

5) Offline training and assessment (key to security)

• IPS/DR: Inverse Propensity Scoring/Doubly Robust on the recommendation logs.
• Replay/Simulators: simulators with custom/provider response models.
• Conservative Regulation: Exit penalty for supporting these logs (CQL/IQL).
• Logger policy: log the probability of impressions (propensity) so that there is a correct estimate.

python value_dr = np.mean(w_ips(r - q_hat) + v_hat) # w_ips = π(a    s)/μ(a    s)

6) Contextual Bandits: Fast Start

• Thompson Sampling (logit): posterior by coefficients → choice of action.
• UCB: for tight budgets and strong sars.
• Regularization fairness/RG: mask unacceptable actions, limit the frequency of impressions.

python β ~ Posterior()         # выбор из апостериорного распределения scores = X @ β         # контекстные оценки actions = top_k(scores, k=slate_size, mask=policy_mask)

7) Slate-RL (kit recommendations)

Goal: to maximize the reward of the entire set (taking into account positions, card competition).
Methods: Listwise-bandits, slate-Q, policy gradient with factorization (Plackett-Luce).
Position correction: propensity by position; randomization within acceptable bounds.

8) Safety, RG/AML and Compliance

• Hard constraints: prohibition of toxic offers, frequency restrictions, "cooling."
• Policy Shielding: filter the action by the RG/AML policy before and after the inference.
• Dual optimization: Lagrange multiplier for restrictions (for example, the share of "aggressive" offers ≤ θ).
• Ethics and fair-use: exclude proxy features, influence audit.

python a = policy.sample(s)
if not passes_guardrails(a, s):
a = safe_fallback(s) # правило/минимальный оффер

9) Data and Serving Architecture

Offline loop

Lakehouse: logs of impressions/clicks/conversions, propensities, cost.
Feature Store (offline): point-in-time features, correct labels.

Training: offline RL (CQL/IQL) + simulators; IPS/DR validation

Online/near-real-time

Feechee: Quick windows (1-60 min), player/session signs, limits and RG/AML masks.
Serving: gRPC/REST, p95 50-150 ms (personalization), canary routing.
Logs: save 'policy _ id', 'propensity', 'slate', 'guard _ mask', actual outcome.

10) Metrics and experimentation

Offline: DR/IPS-assessment value, coverage support, divergence from the logger.
Online: increment to Net Revenue/LTV, RG signals (time-to-intervene), abuse-rate, CTR/CVR/retention.
Risk metrics: CVaR, proportion of guardrails violations, frequency of RG interventions.
Experiments: A/B/n with traffic capping and "kill-switch," sequential-testing.

11) Cost engineering and performance

Complexity of actions: we limit the size of the slate/space of offers.
Cache of features/solutions: short TTLs for popular states.
Decomposition: two-stage (candidate generation → re-rank).
Offline training on schedule: daily/weekly retrain; online - only easy adaptation (bandits).

12) Examples (fragments)

python for rollout in rollouts:
A = advantage(rollout, value_fn)
loss_policy  = -E[ clip_ratio(pi, old_pi) A ]
loss_value  = mse(V(s), R)
loss_safety  = λ relu(safety_metric - safety_cap)
loss_total  = loss_policy + c1loss_value + loss_safety + c2entropy(pi)
step_optimizer(loss_total)

python loss_cql = mse(Q(s,a), target) + α (E_{a~π}[Q(s,a)] - E_{a~D}[Q(s,a)])

python scores = model(x)           # предсказанные полезности scores[~allowed_mask] = -inf     # запретные действия a = argmax(scores) if rand()>eps else random_allowed()

13) Processes, Roles and RACI

R (Responsible): Data Science (RL models/bandits), MLOps (platform/logging/evaluation), Data Eng (features/simulators).
A (Accountable): Head of Data / CDO.
C (Consulted): Compliance/DPO (RG/AML/PII), Legal (terms of offers), Security (secrets/KMS), SRE (SLO/value), Product.
I (Informed): Marketing/CRM, Operations, Support.

14) Implementation Roadmap

1. Context bandit for choosing 1 offer with RG/AML masks and propensity logging.

2. Offline IPS/DR score, limited to A/B (5-10% of traffic), kill-switch.

3. Dashboards: value (DR), CTR/CVR, RG metrics, guardrails violations.

• Slate bandit (N = 3-5 cards), positional corrections; two-stage candidate→re-rank.
• Offline RL (IQL/CQL) with simulator; regular retrain.
• Constrained-RL: limits on aggressiveness/frequency, dual optimization.

• RG intervention (safe RL) policies with strict caps and audits.
• Budget-pacing and bidding (auctions) with CVaR restrictions.
• Interregional adaptation, chargeback at the cost of inference and offers.

15) Pre-sale checklist

• Logs contain 'policy _ id', 'propensity', masks/constraints, outcomes.
• DR/IPS score stable; sufficient data support (overlap with logger).
• Guardrails: inhibit lists, frequency limits, cooldown, kill-switch.
• RG/AML/Legal agreed on rules; audit enabled (WORM for cases).
• Canary release and traffic limits; monitoring value/RG/abuse.
• Award and risk documentation; policy card (owner, version, SLA).
• Cost under control: latency p95, cost/request, slot size, cache.

16) Anti-patterns

Online explor without protection and offline assessment.
Click/bet award excluding abuse and RG → toxic policy.
Lack of propriety and correct causal evaluation by logs.
Too much action space, no masks/capping.
Mixing regions/jurisdictions without residency and rules.
Absence of kill-switch and canaries.

17) The bottom line

RL gives the iGaming platform adaptive policies that maximize long-term value while complying with RG/AML/Legal. The key to safe implementation is offline/conservative methods, correct causal assessment (IPS/DR), strict guardrails, transparent reward, MLOps discipline and gradual rollout. This way you get Net Revenue/LTV growth without compromising on responsibility and compliance.