Tallennettuna:
| Päätekijät: | , |
|---|---|
| Aineistotyyppi: | Recurso digital |
| Kieli: | |
| Julkaistu: |
Zenodo
2026
|
| Aiheet: | |
| Linkit: | https://doi.org/10.5281/zenodo.19618215 |
| Tagit: |
Lisää tagi
Ei tageja, Lisää ensimmäinen tagi!
|
Sisällysluettelo:
- <p>Classical mechanism design theory employs the revelation principle as its core<br>tool, and its validity strictly depends on the fixedness of participants’ type spaces<br>and the validity of the common prior assumption. However, in structurally open<br>multi-agent systems where cognitive architectures can evolve endogenously, partic<br>ipants not only choose actions but may also create new concepts and revise their<br>internal representational frameworks during interaction. Under such conditions,<br>classical mechanism design faces a twofold challenge: participants may not yet<br>have formed a stable conceptualization of their own preferences, and the mecha<br>nism designer cannot exhaustively pre-specify all possible type spaces that may<br>emerge. This paper proposes the post-mechanism design framework, which trans<br>forms the designer’s role from that of a central planner into that of a manager of a<br>cognitive ecosystem. The core task is to guide the multi-agent population, through<br>rule design, toward functional cognitive differentiation rather than pathological<br>polarization. This paper puts forward three principles of post-mechanism design:<br>the concept anchoring protocol compels agents to cross-validate newly introduced<br>concepts against public feedback; the diversity regularizer introduces temporary in<br>centives for cognitive architectural uniqueness into the meta-reward function; and<br>the pathological pruning mechanism identifies and freezes self-referential recursive<br>loops detached from physical feedback based on formal verification tools. The<br>paper further demonstrates how to implement these principles in a multi-agent re<br>inforcement learning environment, enabling the agent population to spontaneously<br>evolve a cognitive division of labor into theorists and experimentalists. This frame<br>work provides a systematic methodology, with both theoretical foundations and<br>engineering operability, for designing AI multi-agent systems endowed with open<br>ended adaptability.</p>