Belief-Desire-Intention (BDI) Architecture

The BDI model is fundamentally defined by three interconnected mental states that drive an agent's decision-making loop.

• Beliefs: The agent's knowledge or information about the world, which may be incomplete or incorrect. This is its internal model.
• Desires: The agent's objectives or goals, representing the motivational state. Desires can be conflicting and are not necessarily consistent.
• Intentions: The subset of desires the agent has committed to pursuing. Intentions are active goals that drive the agent's current and future actions, creating a form of persistence.

This triad forms a practical reasoning cycle: the agent deliberates on its beliefs and desires to form intentions, then uses means-ends reasoning to find plans to achieve those intentions.

BDI agents operate through a continuous, event-driven loop that mimics human practical reasoning. This cycle consists of four primary stages:

1. Belief Revision: The agent updates its belief set based on new percepts from the environment or internal messages.
2. Option Generation: Based on current beliefs and existing intentions, the agent generates a set of possible desires or goals (options) that are achievable.
3. Filtering (Deliberation): The agent selects which options to commit to, transforming them into new intentions. This involves resolving conflicts between competing goals and considering the feasibility of plans.
4. Plan Execution: The agent selects or generates a plan for its active intentions and executes the corresponding actions. It monitors for success, failure, or changes in relevance.

This loop allows the agent to be reactive to environmental changes while maintaining proactive goal-directed behavior.

Unlike models that generate plans from first principles, classic BDI agents typically rely on a plan library—a predefined set of recipes for achieving goals.

• Each plan is a procedural template triggered by an intention and relevant belief conditions.
• Plans can include sub-goals, leading to hierarchical decomposition of tasks.
• Means-ends reasoning is the process of selecting an appropriate plan from the library that satisfies the current intention given the agent's beliefs.

This approach emphasizes efficiency and domain-specific knowledge over universal problem-solving, making BDI suitable for complex, time-constrained environments where pre-defined strategies are known. The agent's adaptability comes from choosing which pre-existing plan to use, not from constructing entirely new ones from scratch.

A defining feature of BDI is the treatment of intentions as commitments. This is not a blind commitment; it is governed by commitment strategies that determine when an agent should drop an intention.

Common strategies include:

• Blind Commitment: Persist with an intention until it is believed to be achieved.
• Single-Minded Commitment: Persist until it is believed either achieved or impossible.
• Open-Minded Commitment: Persist until it is believed achieved, impossible, or irrelevant (e.g., the goal is no longer desired).

This commitment creates persistence, preventing the agent from being overly fickle. It also introduces focus, as resources are dedicated to active intentions. The choice of strategy balances reactivity with tenacity, a key design decision in BDI agent implementation.

BDI agents are inherently event-driven. Changes in the environment (external events) or the internal state (goal achievement/failure, message receipt) trigger the reasoning cycle.

• Event Queue: Incoming events are placed in a queue for processing, allowing the agent to handle multiple, potentially concurrent, stimuli.
• Concurrent Intentions: An agent can actively pursue multiple intentions simultaneously. This requires managing shared resources and interleaving the execution steps of different plans.
• Interruption and Resumption: More important events can interrupt the execution of a plan for a current intention. The agent may later resume or replan.

This architecture supports the design of sophisticated agents that can manage multiple goals, respond to urgent situations, and maintain ongoing activities—a crucial capability for real-world deployment.

The BDI model has strong roots in philosophy of mind (Bratman's theory of human practical reasoning) and formal logic.

• Logical Formalizations: The concepts of belief, desire, and intention have been formalized using modal logics (e.g., BDI logics) to enable precise specification and verification of agent properties.
• Reference Implementations: The model has been realized in several influential agent programming platforms and frameworks.
  • PRS (Procedural Reasoning System): An early seminal implementation.
  • JACK™ Intelligent Agents: A commercial Java-based framework.
  • Jason: A widely-used open-source interpreter for an extended version of AgentSpeak(L), a BDI-inspired programming language.
  • Jadex: A BDI reasoning engine integrated into the JADE multi-agent middleware.

These implementations provide the tools to build BDI-based agents, handling the reasoning cycle, plan management, and event processing.

A formal theory extending individual BDI reasoning to collaborative teams. It models when a group of agents has a mutual commitment to a shared goal. Key principles include:

• Joint Persistent Goal: The team commits until they mutually believe the goal is achieved, impossible, or irrelevant.
• Mutual Belief: Agents not only intend the goal but also believe that others intend it and believe in the mutual commitment.
• This theory provides a rigorous logical foundation for teamwork, ensuring agents coordinate reliably on collective tasks, unlike simple parallel individual intentions.

A formal language enabling BDI agents (and others) to exchange messages with unambiguous meaning. It provides the "speech acts" through which agents share beliefs, request actions (influencing desires), and commit to plans (establishing intentions).

• FIPA ACL is the most prominent standard, defining communicative acts like inform, request, cfp (call for proposal), and propose.
• Each message has precise semantics defined by pre- and post-conditions on the sender's and receiver's mental states, enabling reliable interaction between heterogeneous agents.

A normative framework that creates explicit obligations between agents, providing a mechanism for trust and coordination that complements internal BDI reasoning. A commitment is a tuple: Commit(Debtor, Creditor, Condition).

• Creates a persistent social relationship, obliging the debtor to bring about the condition for the creditor.
• Differs from a private intention; it is a public, verifiable promise that other agents can rely on for planning.
• Used to model contracts, protocols, and institutional rules, adding a layer of social accountability to agent interactions.

A planning methodology often integrated with BDI architectures to manage the complexity of plan generation. An HTN planner works by recursively decomposing high-level tasks into subtasks via pre-defined methods.

• Task Library: Contains abstract tasks (e.g., DeliverPackage) and primitive tasks (e.g., NavigateTo).
• Decomposition Methods: Rules for breaking down abstract tasks into networks of subtasks.
• In a BDI context, the intention structure often manages the execution of a plan (task network) generated by an HTN planner, linking high-level desires to executable actions.

A coordination approach where agents exchange and merge their local plans to form a partial view of a cooperative global plan. It addresses the limitation of purely local BDI reasoning in multi-agent settings.

• Agents broadcast their local plans (intentions).
• They analyze received plans to detect interactions (conflicts, synergies, dependencies).
• They then modify their own local plans to resolve conflicts or exploit cooperation opportunities.
• This creates a decentralized, evolving coordination structure without a central controller, suitable for dynamic environments.

A framework for governing open multi-agent systems by defining the norms, rules, and interaction spaces within which autonomous agents (including BDI agents) operate. It provides the "social infrastructure" for coordination.

• Scene Protocols: Define allowed conversation flows within a virtual room (e.g., a Dutch auction room).
• Normative Rules: Specify obligations, permissions, and prohibitions for agents in roles.
• Sanctions: Enforce rules through rewards or penalties.
• This external, institutional layer constrains and guides the practical reasoning of individual BDI agents, ensuring societal order.