3.3 Advanced AI Engine

Generative AI(GenAI) combined with Reinforcement Learning(RL) will bring new possibilities to the content generation and intelligent agent ecosystem of the IAW games.

GenAI research has rapidly advanced AI content generation to the production stage, with chatGPT being one of the most popular platforms. Using GenAI can lower the barrier to entry for game design and maximize the potential for autonomous evolution of game maps and characters.

Apart from chatGPT, Midjourney could directly generate images from the prompt text, and Runway AI already has the computational ability to convert text into short videos. If specific prompts are generated based on the parameter server and the player's background, the game scenes and content can be dynamically updated in real time, which can definitely realize the production of game autonomous NPCs to a certain extent.

In addition to GenAI-powered machine content creation, RL AI exploration is another key technique for building an autonomous AI agent ecosystem.

Recently, researchers from Stanford University and Google built a 2D virtual game environment called Smallville - Generative Agents: Interactive Simulacra of Human Behavior [7]. They placed 25 AI agents based on ChatGPT in the environment for training. The study found that the 25 AI agents achieved a credible simulation of human behavior. They could not only communicate with each other, but also interact with their surroundings, remember and recall what they had done and observed, and make appropriate decisions.

The study's findings suggest that RL AI has the potential to be used to create more realistic and engaging AI agents.

This paper proposes a three-layer architecture for the IAW world: Data Layer, Model Layer, and Agent Layer.

The Data Layer works with the IAW engine framework to unify OnChain and OffChain state management. The Model Layer fine-tunes open-source large models and combines them with the AI token settlement system to form a diversified dedicated model market. Finally, based on the Model Layer, a large number of agents (GenAI and RL AI) are created to form the machine characters in the IAW world.