Project ALJV is a 2-team shooter simulation prototype built in Unreal Engine 5. The project focuses on emergent combat behavior by utilizing the Learning Agents plugin to train AI bots dynamically via Reinforcement Learning.
- Engine: Unreal Engine
5.7.4 - AI & Machine Learning: Learning Agents Plugin
- Algorithm: Proximal Policy Optimization (PPO)
- Scripting: Blueprints (Visual Scripting)
- Multi-Agent Combat Simulation: Features an active 8v8 battleground (Team Red vs. Team Blue), where all 16 robotic agents are driven simultaneously by reinforcement learning models.
- Asymmetric Observation Logic: The teams evaluate the environment differently to simulate varied tactical approaches:
- Team Red: Focuses its observations on prioritizing and engaging the closest visible enemy.
- Team Blue: Takes into account the spatial positioning and layout of all active enemies on the field.
- Custom Reward Function: Agents optimize their behavior based on a precise reward/penalty system. They receive positive rewards for successful eliminations (calculated via accurate line-trace intersections with enemies) and negative rewards (penalties) for friendly fire.
- Continuous Runtime Learning: Bots do not just rely on pre-trained models; they learn and adapt their policies organically over continuous epochs as the simulation plays out.
- Real-Time Scoreboard: Features a live UI system to track team performance, specifically kills.
This project was created by:
- Codarcea Alexandru-Christian
- Velișan George-Daniel