Reducing On-Chain Costs

• We explored various solutions to reduce the on-chain costs of the game.
• Initially, we considered using circom to write circuits and generate SNARK proofs. However, this approach proved to be too limited in scalability, supporting only simple game logic.
• Another idea was to write the entire game in Solidity and put it in a rollup. However, this approach posed challenges in rendering the game frame to the frontend for display, especially without emitting all data in events, which incurred high costs.
• Ultimately, we opted to use a ZK coprocessor to provide proofs of the computational process, publishing the ZKP on-chain for verification. Additionally, we encoded the game board as unit256, with each end bit of the canvas not displaying, significantly reducing costs.

AI Model Selection

• We conducted extensive trials in selecting the appropriate AI model for the game.
• Initially, we explored reinforcement learning approaches like Q-learning but found them impractical due to the game's complexity, and limited time available, leading to high training costs. Building such models also requires non-trivial heuristic fine-tuning, which is not practical for the hackathon
• We ultimately adopted the Minimax search algorithm with limited search depth, which seamlessly integrates with existing technology frameworks and provides relatively fast feedback on gameplay experience.