An interactive reinforcement learning system for LiveOps decisioning that uses Q-learning, safety-gated recommendations, rollout forecasting, benchmark comparisons, and LangChain RAG explanations to simulate player intervention strategies.
LiveOps RL Decision Engine is an interactive machine learning demo for game LiveOps decisioning. The project uses a safety-gated reinforcement learning policy to recommend player interventions, simulate match outcomes, forecast rollout impact, and explain decisions through a lightweight RAG assistant.
In this walkthrough, I show the full system in action: scenario presets, manual player-state controls, live win/frustration/churn-risk metrics, safety-gate behavior, benchmark comparisons, rollout forecasting, policy/audit panels, and the agent chat experience.
Tech stack:
Python, FastAPI, Uvicorn, PyTorch, Docker, Hugging Face Spaces, LangChain, Gemini, Ollama, Google Cloud, and BigQuery.
Project features:
– Tabular Q-learning policy for LiveOps recommendations
– Deterministic safety gate for policy-compliant actions
– Browser-based game arena simulation
– Rollout forecasting over multiple match cycles
– Benchmark comparison against baseline strategies
– Lightweight LangChain RAG assistant for policy and decision explanations
– Optional GCP, BigQuery, and Gemini cloud path
– Docker deployment on Hugging Face Spaces