Immitation Learning in Football
Creating defensive player agents to imitate real-world tactics
Joint Policy Agent Inference Demo
Single Policy Agent Inference Demo
During my exploration of data-driven football analytics, I saw the potential in developing intelligent agents within simulated football environments. This idea was echoed in a competition hosted by Manchester City, which focused on training AI agents in simulated football settings (see competition). Their approach on training agents in synthetic environments like those found in this open-source simulation was undeniably exciting and advanced.
However, looking at this from the perspective of actual football development, I recognized a few limitations:
- Simulations can't capture the full complexity of real football matches due to imperfect physics and limited tactical representation.
- Agents trained purely in synthetic environments often have limited applicability in real-world coaching or player development.
- Such agents are often optimized more for "gaming" the simulation than mirroring real football player behavior.
Motivated by these observations, I sought a more grounded approach—one that could make a tangible impact on real-world coaching. I came across the paper Coordinated Multi-Agent Imitation Learning, which uses actual match position data to imitate real player movement. When I discussed this approach with our football SMEs (Subject Matter Experts), they agreed it showed promising potential.
The core idea is to replicate coaching behavior: helping players improve by analyzing and correcting their movement—something traditionally done in film rooms or training grounds. I implemented an adapted version of the imitation learning algorithm using a curated dataset of real matches. The goal is that, as more data becomes available, the system could actively support coaches in guiding players based on real-world performance insights.