Tsukada Lab. Computer Network Group

Deliberation Trace-Based Commitment Control for Multi-Agent Systems under Partial Observability

Hyeonbin Kang, Wencan Mao, Manabu Tsukada

International Conference on Intelligent Computing and its Emerging Applications (ICEA 2026), Kaohsiung, Taiwan, 2026年11月19日

3行要約

  • D-TRACEは部分観測の複数エージェントに対し、追加照会、決定、安全な代替行動への切り替えを判断する。
  • 熟慮の履歴を時系列モデルで処理し、次の照会に要するコストと得られる改善を比較する。
  • 現在の特徴が同一の評価ケースでは、履歴を使うことで状況の違いを識別できた。

アブストラクト(原文)

Multi-agent systems under partial observability must integrate specialized-agent judgments while deciding whether to acquire more information or act. Fixed or uncertainty-based stopping criteria may not distinguish valuable deliberation from repetition after information value is exhausted. D-TRACE integrates planning, risk assessment, memory, reflection, and goal evaluation through a Shared Cognitive Workspace and hierarchical coordination while recording prior query outcomes, information gains, agent judgments, and candidate-action ranking changes. It continues QUERY only when the expected benefit of a query-induced action change exceeds query and query-count-dependent delay costs, then applies a risk constraint to select COMMIT or SAFE FALLBACK. Numerical analysis shows that the query boundary adapts to expected improvement, query cost, and prior query count rather than a fixed uncertainty threshold.

引用方法

H. Kang, W. Mao, and M. Tsukada, “Deliberation Trace-Based Commitment Control for Multi-Agent Systems under Partial Observability,” in International Conference on Intelligent Computing and its Emerging Applications (ICEA 2026), Kaohsiung, Taiwan, 2026.

BibTeX

@inproceedings{Kang2026,
  title         = {Deliberation Trace-Based Commitment Control for Multi-Agent Systems under Partial Observability},
  author        = {Hyeonbin Kang and Wencan Mao and Manabu Tsukada},
  year          = {2026},
  date          = {2026-11-19},
  urldate       = {2026-11-19},
  booktitle     = {International Conference on Intelligent Computing and its Emerging Applications (ICEA 2026)},
  address       = {Kaohsiung, Taiwan},
  abstract      = {Multi-agent systems under partial observability must integrate specialized-agent judgments while deciding whether to acquire more information or act. Fixed or uncertainty-based stopping criteria may not distinguish valuable deliberation from repetition after information value is exhausted. D-TRACE integrates planning, risk assessment, memory, reflection, and goal evaluation through a Shared Cognitive Workspace and hierarchical coordination while recording prior query outcomes, information gains, agent judgments, and candidate-action ranking changes. It continues QUERY only when the expected benefit of a query-induced action change exceeds query and query-count-dependent delay costs, then applies a risk constraint to select COMMIT or SAFE FALLBACK. Numerical analysis shows that the query boundary adapts to expected improvement, query cost, and prior query count rather than a fixed uncertainty threshold.},
  keywords      = {Other},
  pubstate      = {published},
  tppubtype     = {inproceedings}
}