Simon Hedrich
Simon Hedrich is a computer scientist and AI enthusiast currently completing his Master’s degree in Computer Science. His academic and professional journey is marked by a deep interest in bridging the gap between theoretical research and practical AI engineering.
Through his work at inovex GmbH, Simon has demonstrated expertise in specialized areas of Artificial Intelligence, including computer vision and the use of synthetic data to enhance small object detection. His technical writing highlights his ability to leverage generative AI models, such as Stable Diffusion, to solve complex real-world challenges like training data scarcity.
Session
What does it take to build an AI that learns to play strategy games from scratch? Over the past year, I chose to explore this question out of personal fascination with game AI — as a seminar project for college, but really as a hobby. The result was a complete reinforcement learning environment for Antiyoy, a turn-based strategy game played on hexagonal grids.
The journey raised intriguing challenges: How do you represent hexagonal game boards for neural networks? What do you do when your AI has over 4,000 possible actions to choose from? How do you design rewards that teach strategy rather than just reward flailing in the right direction? This talk shares how these problems were approached using Python's modern ML ecosystem—Gymnasium, PyTorch, and PPO training—ultimately producing an agent that wins nine out of ten games against a random opponent. Whether that qualifies as "strategic play" is a question the agent and I still disagree on.
Whether you're curious about building custom RL environments, interested in game AI, or just wondering what reinforcement learning actually looks like when it half-works, you'll leave with practical insights and a healthy dose of realistic expectations.