Belief-Desire-Intention (BDI) Model

Intentions are committed plans—desires that the agent has chosen to actively pursue and has adopted a plan to achieve. They represent a promise to act.

• Function: Focus resources, drive planning, and ensure persistent goal-directed behavior.
• Key Property: Intentions create a filter of admissibility for new options, promoting stability. An agent will typically not adopt a new intention that conflicts with an existing one.
• Example: From its desires, an agent may form the intention deliver_package(P123, Address_456) and commit to a specific route plan.

The BDI agent operates via a continuous practical reasoning loop that mediates between its components. This loop has two main phases:

1. Deliberation (Goal Selection): The agent reviews its beliefs and desires to decide which goals to commit to as intentions. This often involves filtering options based on relevance and feasibility.

• Means-Ends Reasoning (Plan Selection): For each active intention, the agent retrieves or generates a plan—a sequence of actions—from a plan library that is expected to achieve the goal given current beliefs.
• Execution & Monitoring: The agent executes the first step of the selected plan, then monitors the world (updating beliefs) to see if the plan is still viable. If a plan fails, the agent replans or reconsider its intentions.

A BDI agent does not typically plan from first principles. Instead, it relies on a plan library—a predefined set of recipes or procedures that map goals to action sequences under certain context conditions.

• Structure: A plan is often represented as Goal : Context <- Body. The Body is the action sequence, executed only if the Context (a set of beliefs) is true.
• Function: Enables fast, reactive behavior by using pre-compiled knowledge.
• Example: A plan for charge_battery might have a context battery_below(20%) and a body [navigate_to(charging_station), connect_charger, wait_until(battery_above(95%))].

BDI agents are inherently event-driven. The reasoning loop is triggered by changes in:

• Belief Updates: New perceptual input or communicated information.
• Goal Addition: New desires posted internally or received from a user/other agent.
• Plan Failure: The context of a running plan becomes false or an action fails.

This architecture allows agents to be reactive (responding promptly to changes) while remaining goal-oriented. For instance, a new belief road_blocked(main_street) would trigger a replanning event for any intention using that route.

Recursive modeling is a computational approach where an agent models not only the world but also the mental models of other agents, potentially nesting these models to multiple levels (e.g., 'I think that you think that I think...').

• Strategic Depth: Essential for complex negotiation, poker, and any adversarial or cooperative scenario requiring anticipation of an opponent's moves.
• BDI Extension: A BDI agent can be equipped with recursive modeling to maintain beliefs about other agents' beliefs, desires, and intentions.
• Computational Challenge: The complexity grows exponentially with recursion depth, often requiring bounded rationality approximations.

Inverse planning (or Bayesian inverse reinforcement learning) is a formal, probabilistic framework for inferring an agent's hidden goals and beliefs by reasoning backwards from its observed actions, under the assumption that the agent is approximately rational.

• Mathematical Foundation: Uses Bayesian inference to compute the posterior probability of goals given actions and a model of planning (like BDI).
• Precision: Provides a rigorous alternative to heuristic plan recognition, quantifying uncertainty over possible mental states.
• Use Case: Critical for building AI that can collaborate seamlessly with humans by continuously inferring their changing objectives.

Multi-agent epistemic logic is a formal logical system used to rigorously reason about the knowledge and beliefs of multiple interacting agents. It provides the mathematical syntax and semantics for expressing statements like 'Alice knows that Bob does not know the secret.'

• Formalizes Mental States: Provides tools to define belief (as in BDI), knowledge, and common knowledge with precise, unambiguous meaning.
• Enables Verification: Allows system designers to prove properties about what agents can or cannot deduce about each other's states.
• Foundation: Serves as the theoretical backbone for designing and analyzing communication protocols and cooperative algorithms in distributed AI systems.

Shared mental models are overlapping or aligned internal representations of a task, team, or situation held by members of a group. They enable coordinated, efficient action without the need for constant, explicit communication.

• Team Coordination: In a multi-agent BDI system, agents may develop shared mental models about the environment, team roles, and procedures.
• Reduces Overhead: Facilitates implicit coordination, as agents can accurately predict what their teammates will do based on a shared understanding of the plan.
• Engineering Challenge: Requires mechanisms for belief alignment, such as specific communication acts or joint experiences, to establish and maintain the shared model.