| 研究課題/領域番号 |
23K28145
|
|---|
| 補助金の研究課題番号 |
23H03455 (2023)
|
|---|
| 研究種目 |
基盤研究(B)
|
| 配分区分 | 基金 (2024)
補助金 (2023) |
|---|
| 応募区分 | 一般 |
|---|
| 審査区分 |
小区分61030:知能情報学関連
小区分60030:統計科学関連
合同審査対象区分:小区分60030:統計科学関連、小区分61030:知能情報学関連
|
|---|
| 研究機関 | 京都大学 |
|---|
研究代表者 |
Rafik Hadfi 京都大学, 情報学研究科, 特定准教授 (30867495)
|
|---|
| 研究期間 (年度) |
2024-04-01 – 2026-03-31
|
| 研究課題ステータス |
交付 (2024年度)
|
| 配分額 *注記 |
14,170千円 (直接経費: 10,900千円、間接経費: 3,270千円)
2025年度: 7,020千円 (直接経費: 5,400千円、間接経費: 1,620千円)
2024年度: 4,680千円 (直接経費: 3,600千円、間接経費: 1,080千円)
2023年度: 2,470千円 (直接経費: 1,900千円、間接経費: 570千円)
|
|---|
| キーワード | AI / Agents / Autonomy / Agency / Ethics / Morality / Game Theory / Psychology / Artificial Intelligence / Trustworthy AI / Trust |
|---|
| 研究開始時の研究の概要 |
The respect for human autonomy is a crucial principle in developing trustworthy AI. This project investigates autonomy, explores the factors influencing it, and tests it with human and AI agents.
|
| 研究実績の概要 |
The project has progressed through a series of results addressing ethical autonomy in AI agents. The most recent results include a formal framework to quantify ethical autonomy by integrating Boltzmann rationality with moral philosophy (Joseph Raz's autonomy), offering practical tools to enhance human-centered autonomy in AI, especially in safety-critical domains such as eldercare and assistive robotics. Another notable achievement is developing a theoretical liberal perfectionist framework based on Razian autonomy, advocating for AI systems that actively support human autonomy through rational choice and independence from coercion. Additionally, the project examined moral autonomy as a foundation for ethical AI, proposing methods to embed moral decision-making capabilities into autonomous agents. In the context of objective 3, the research explored how personality traits influence agentic dynamics using LLM-based simulations, demonstrating the behavioral alignment of synthesized personalities with human-like behaviors. Furthermore, a novel multi-observer method for LLM personality assessment was proposed, improving accuracy by employing diverse observer contexts. These results advance ethical autonomy's theoretical and practical foundations in AI.
|
| 現在までの達成度 |
現在までの達成度
2: おおむね順調に進展している
理由
First, on the theoretical scope, progress was achieved by adopting Joseph Raz’s perspective on autonomy from moral philosophy, allowing for a principled understanding of ethical autonomy in AI agents. Second, on the mathematical front, the project successfully quantified autonomy within various game-theoretic settings, employing a Boltzmann rationality model to assess decision-making under bounded rationality and environmental constraints. This formalization provided a robust framework to analyze autonomy profiles in strategic interactions. Finally, in terms of practicality, progress was made by constructing LLM agents to simulate agentic dynamics. These simulations are designed to capture nuanced interactions between agents, forming a foundation for future evaluations of autonomy within complex game-theoretic contexts. These theoretical, mathematical, and practical advances have enabled the project to address the complex challenge of embedding human-centered autonomy in AI systems.
|
| 今後の研究の推進方策 |
Future plans consist of evaluating autonomy in complex economic domains, particularly focusing on negotiation scenarios. Building on previous results, we will expand the autonomy framework to address multiagent negotiations with intricate strategic interactions. This task involves integrating the Raz-Boltzmann model with dynamic decision-making processes in real-world bargaining contexts. The focus will also shift to linguistic domains, where LLM-based simulation results will be assessed using the autonomy framework developed in objectives 1-2. This assessment will help validate how LLM agents exhibit autonomy when interacting linguistically in negotiation settings. The development of additional simulations will leverage the existing platform, making it more scalable. We will also incorporate psychological profiles, allowing agents to exhibit diverse personality traits and strategic behaviors, mirroring human-like behavioral dynamics.
|
