Integrating Machine Learning into Belief-Desire-Intention Agents: Current Advances and Open Challenges

TL;DR

The paper tackles fragmentation in integrating ML with rational BDI agents by delivering a fine-grained, BDI-aligned survey of 98 primary works. It systematically maps ML techniques to Beliefs, Desires, Intentions, and Actions, highlighting trends (notably belief representation and planning with LLMs) and gaps (sensing, belief revision, knowledge enrichment, desire/option modeling, and online learning). The authors identify open challenges and practical directions, such as belief-driven sensing, neurosymbolic belief revision, verifiable plan generation, and the need for online adaptation to avoid static deployment limitations. This work provides a principled foundation for designing robust, trustworthy ML-BDI agents and guides future research toward principled integration rather than eliminativist tendencies.

Abstract

Thanks to the remarkable human-like capabilities of machine learning (ML) models in perceptual and cognitive tasks, frameworks integrating ML within rational agent architectures are gaining traction. Yet, the landscape remains fragmented and incoherent, often focusing on embedding ML into generic agent containers while overlooking the expressive power of rational architectures--such as Belief-Desire-Intention (BDI) agents. This paper presents a fine-grained systematisation of existing approaches, using the BDI paradigm as a reference. Our analysis illustrates the fast-evolving literature on rational agents enhanced by ML, and identifies key research opportunities and open challenges for designing effective rational ML agents.

Figures (14)

• Figure 1: Distribution of works available in the literature over the various BDI modules (\ref{['item:rq2']}). Belief-related papers are analyzed in Section \ref{['sec:ml_beliefs']}, while Section \ref{['sec:ml_desires']} delves into ML and BDI desires. ML approaches targeting intentions are studied in Section \ref{['sec:ml_intentions']}. Finally, in Section \ref{['sec:ml_actions']} we investigate the action-enabling ML papers.
• Figure 2: Distribution of ML models used in the ML-BDI literature across all years.
• Figure 3: Distribution of ML models used in the ML-BDI literature per year.
• Figure 4: Distribution of source code availability of ML-BDI architectures per year.
• Figure 5: Distribution of belief representation classes. See legend of \ref{['tab:belief-taxonomy']}.