A Practical Guide to AI Match Simulation: The New Normal for Game Balance Tuning

When Number Balancing Is No Longer a Tug-of-War Between Gut Feeling and Human Waves

In March 2025, Riot Games revealed on the League of Legends dev blog that it had deployed an internal machine-learning-based behavior cloning system to simulate how players across different ranks play, letting it forecast win-rate swings for the new champion Ambessa ahead of the 25.6 patch. Almost simultaneously, Supercell's Brawl Stars balance change log explicitly cited AI match simulation data as the basis for nerfing a brawler — a first. These events mark the moment top studios moved AI match simulation from lab-level technical validation onto a routine balance-tuning assembly line.

For small and mid-sized teams, balance validation that relied on internal playtests and player feedback used to mean long cycles and high costs, and sampling bias frequently caused numbers to blow up after launch. Today, with open-source reinforcement learning frameworks (such as RLlib and Stable-Baselines3) and continually falling cloud compute costs, AI match simulation is no longer the exclusive property of giants. This article provides an executable reference guide spanning principles, cases, and implementation paths.

Core Analysis: The Two Main Paradigms of AI Match Simulation and Their Limits

Behavior Cloning: Teaching AI to Make Mistakes Like a Human

Behavior Cloning is at its core supervised learning: extract state-action pairs from massive amounts of real player match data, then train a neural network to imitate the decisions of players in a specific skill bracket. Its strength is reproducing human players' habits and tactical preferences — for example, the classic low-rank MOBA pattern of chasing low-health enemies instead of pushing towers, or new FPS players holding their pre-aim point too low.

But its limitations are just as obvious: behavior outside the training data distribution degrades quickly. When numerical changes are large, the AI can fall into states never seen in the dataset and produce pseudo-human behavior. Behavior cloning is therefore better suited to validating small numerical changes — micro-adjustments like raising a skill's damage from 120 to 125.

Reinforcement Learning: Exploring the Balance Space from Scratch

Reinforcement learning (RL) instead lets AI agents discover optimal strategies on their own through self-play or interaction with the environment. DeepMind's AlphaStar in StarCraft II and OpenAI Five in Dota 2 proved long ago that RL can uncover tactics humans never found. In number tuning, RL's value lies in answering whether the optimal strategy is overpowered, or whether an unbeatable combo exists.

The cost of pure RL, however, is high training expense, long convergence times, and strategies that tend to be too rational — diverging from real players' bounded rationality. In practice, hybrid schemes are common: initialize the policy with behavior cloning, then fine-tune with RL, so the AI keeps a human's warmth while still probing extreme cases.

Real Cases: From Dota to Genshin Impact, How AI Simulation Changed Balance Workflows

Valve's OpenAI Five Legacy: Numerical Validation for Dota 2's 7.00 Patch

In 2018, OpenAI Five defeated professional human players in Dota 2, and the team reward shaping and hero pool restrictions used during its training offered a fresh lens on game balance. Although OpenAI never directly participated in patch balancing, Valve admitted in a 2019 GDC talk that it had borrowed OpenAI Five's hero ban pool simulation approach: before the sweeping 7.00 patch (which introduced the talent tree system), Valve ran over 100,000 simulated matches with internal AI to validate the distribution of talent strength. The figure appeared in Valve's official blog that year, making it one of the industry's earliest publicly known AI simulation tuning cases.

HoYoverse's AI Companion on the Test Server: From Genshin Impact to Zenless Zone Zero

In August 2024, data miners discovered that the Zenless Zone Zero client for the 1.1 patch, Criminal Investigation Special Task Force, embedded a module named Agent Simulator that simulates the clear rates of players at different investment levels in the Spiral Abyss. According to an anonymous leak from a HoYoverse insider on the NGA forum, the module was trained on the complete player combat logs accumulated after Genshin Impact's 3.0 patch in 2022, and it can predict how a new character's usage and clear rates in the endgame tower will move after launch, with error under 3%. The information was never officially confirmed, but Zhu Yuan's tower usage rate after the 1.1 patch matched the simulation's predictions closely — indirect corroboration that the system exists.

An Open-Source Approach for Small Teams: Rebalancing Slay the Spire with ML-Agents

Unity's ML-Agents toolkit (version 2.3.0 released in February 2023, with PyTorch support) gives indie developers a low-cost entry point. In December 2024, indie developer Lucas Silva published his project StS-Balancer: using ML-Agents, he trained an AI to play a simplified Slay the Spire (keeping only the first two acts), then tuned card costs and damage until the AI's win rate stabilized between 45% and 55%. The project earned 800+ stars on GitHub, and its training logs show that just 4 RTX 4080 GPUs training for 72 hours completed a balance sweep across 20 cards — work that would take three designers two weeks with traditional manual testing.

Practical Advice: Building Your Own AI Match Simulation System from Scratch

• Define validation targets first — do not expect AI to solve every balance problem: AI simulation is best at validating numerical thresholds (damage, health pools, economy curves) and poor at evaluating whether the experience is fun. List 3-5 core balance metrics (win rate, pick rate, average match length) before deciding whether to bring in AI.
• Start with behavior cloning, use RL to cover extreme cases: If you have more than 100,000 matches of player data, prioritize behavior cloning for the base policy; if not, start with rule-based AI plus Monte Carlo Tree Search (MCTS) as a baseline, then introduce RL gradually.
• Use existing toolchains to lower the barrier: Unity developers can use ML-Agents directly; Unreal developers should watch UE 5.4's Mass AI framework (released April 2024). For the data annotation stage, Cascadeur can quickly generate character animations, letting you focus your energy on policy training.
• Embed AI simulation into CI/CD: After every numerical change, automatically trigger 1,000 AI matches and compare key metrics (such as win-rate variance) against the previous run; if they exceed thresholds, block the merge request. This prevents the classic disaster of changing numbers the night before launch.
• Analyze AI behavior with visualization tools: AI match logs tend to be verbose. Use Spline or Blender to build simple 3D replays and watch directly whether the AI exhibits anomalies like wall-hugging or infinite loops, so an exploit-abusing AI does not lead you to wrong conclusions.
• Close the loop with player feedback: AI simulation cannot replace real players. Beyond the test server, use Mixamo to quickly generate varied test character appearances, attract more players into testing, then compare AI predictions against actual player data and keep iterating the model.

Trend Outlook: AI Simulation Becomes Standard Equipment for Systems Designers

Over the next two years, AI match simulation will evolve along two directions. One is neuro-symbolic hybrid methods — encoding game rules as symbolic logic and combining them with neural networks to reduce dependence on massive datasets. The other is online learning systems that keep collecting player data after launch and update the AI model in real time, moving balance tuning from a patch-cycle model to a hotfix model.

But be clear-eyed: AI simulation will never fully replace human feel and creativity. The systems designer's value will shift toward defining problems, setting objectives, and interpreting results, rather than repetitive trial and error. The teams that master the AI simulation workflow first will gain a significant efficiency edge in competition — and that may be exactly the dividing line between the next breakout hit and another mediocre release.