curiositech/rao-georgeff-1991-modeling-rational-agents-bdi

SKILL: BDI Agent Architecture

Name: BDI Rational Agent Design
Description: Formal framework for modeling intelligent agents through Beliefs, Desires/Goals, and Intentions as distinct mental states with rigorous semantics
Author/Source: Anand S. Rao & Michael P. Georgeff (1991)
Activation triggers: agent architecture, intelligent systems, commitment reasoning, intention modeling, action planning, side-effects problem, goal decomposition, multi-agent systems, autonomous behavior

DECISION POINTS

Primary Decision Tree: Commitment Strategy Selection

Environment Stability Assessment:
├─ STABLE environment + HIGH resources
│  └─ Choose BLIND commitment (AI₁)
│     • Persist until believed achieved
│     • Maximize goal completion
├─ UNSTABLE environment + LIMITED resources  
│  └─ Choose SINGLE-MINDED commitment (AI₂)
│     • Drop if impossible OR achieved
│     • Balance persistence with realism
└─ HIGHLY DYNAMIC + EXPLORATORY goals
   └─ Choose OPEN-MINDED commitment (AI₃)
      • Drop if no longer desired OR achieved
      • Track changing preferences

Goal Decomposition Decision Tree:
├─ PRIMITIVE action (under agent control)
│  └─ INTEND(does(action)) → guarantees execution
├─ ACHIEVEMENT goal (world determines outcome)
│  └─ INTEND(achievement) → guarantees attempt only
└─ COMPLEX goal requiring subgoals
   ├─ IF sequential: INTEND(q) before INTEND(p) where q enables p
   └─ IF disjunctive: Choose specific branch to commit to

Side-Effects Resolution:
├─ BELIEVE(action → wanted_effect AND unwanted_effect)
│  └─ Can INTEND(does(action)) without INTEND(unwanted_effect)
│     • Intention-worlds select specific branches
│     • Avoid closure under believed implications
└─ Multiple paths to same goal
   └─ Select path minimizing unwanted side-effects

Multi-Agent Coordination Decision Tree

Other Agent's Mental States:
├─ What I BELIEVE they INTEND
│  └─ Plan coordination assuming their commitment
├─ What I INTEND regarding their actions
│  └─ My commitment to outcomes involving them
└─ Commitment conflicts detected
   ├─ IF my commitment is stronger → maintain, negotiate
   └─ IF their commitment is stronger → revise, delegate

FAILURE MODES

1. Intention-Goal Collapse

Detection: If you find yourself saying "intention = persistent goal" or treating INTEND(p) as equivalent to GOAL(p) + high_priority.

Diagnosis: Missing the independence of mental attitudes. Intentions have separate persistence conditions from goals.

Fix: Implement separate axioms for each attitude. An agent can intend something while no longer desiring it (obligation case) or desire something while not committing to it (wish case).

2. Logical Closure Trap

Detection: If BELIEVE(p → q) AND INTEND(p) forces INTEND(q), creating unwanted commitments to side-effects.

Diagnosis: Treating intentions as closed under logical entailment instead of using sub-world compatibility.

Fix: Apply geometric constraints: intention-worlds are sub-worlds of belief-worlds. You intend specific branches, not entire logical consequences.

3. Wrong Commitment Strategy

Detection: Agent gives up too easily (needed blind commitment) or persists irrationally (needed open-minded commitment).

Diagnosis: Mismatched persistence axiom for environmental demands.

Fix: Reassess environment stability and resource constraints. Switch axioms: AI₁ for stable/high-resource, AI₂ for collaborative/resource-limited, AI₃ for dynamic/exploratory.

4. Action-Result Confusion

Detection: Treating failed outcomes as defective intentions when environmental factors prevented success.

Diagnosis: Confusing volitional commitment (does(action)) with result achievement (succeeds(action)).

Fix: Apply AI₄ correctly: INTEND(does(e)) guarantees execution, but INTEND(achieves(goal)) only guarantees attempt.

5. Branch-World Conflation

Detection: Using branching time to represent epistemic uncertainty instead of choice options.

Diagnosis: Collapsing two-dimensional structure (choices within worlds, uncertainty across worlds).

Fix: Separate optional/inevitable (branch quantification within worlds) from BEL/GOAL/INTEND (world quantification across epistemic alternatives).

WORKED EXAMPLES

Example 1: Multi-Agent Coordination with Commitment Revision

Scenario: Two autonomous robots (A and B) must coordinate to move a heavy table. Initially, both have GOAL(table_moved) and different movement strategies.

Initial State:

• Robot A: BELIEVE(I_can_push_alone), GOAL(table_moved), INTEND(push_from_north)
• Robot B: BELIEVE(need_coordination), GOAL(table_moved), INTEND(coordinate_with_A)

Decision Point Navigation:

1. Environment Assessment: Dynamic (robots moving), limited resources (both needed)
2. Commitment Strategy: Single-minded (AI₂) - drop if impossible, maintain if achievable
3. New Information: A discovers table is heavier than expected - BELIEVE(¬I_can_push_alone)

Expert vs. Novice:

• Novice: Drops INTEND(push_from_north), starts new planning from scratch
• Expert: Maintains GOAL(table_moved), revises strategy to INTEND(coordinate_with_B), leverages B's existing coordination intention

Resolution: A adopts compatible sub-goal INTEND(push_while_B_pulls), maintaining higher-level commitment while adapting method.

Example 2: Side-Effect Handling in Medical Diagnosis

Scenario: Diagnostic AI system must recommend treatment knowing it causes side-effects.

Setup:

• BELIEVE(chemotherapy → tumor_shrinkage AND nausea AND fatigue)
• GOAL(tumor_shrinkage)
• Patient explicitly states: NOT GOAL(nausea)

Decision Tree Navigation:

Treatment recommendation:
├─ Can INTEND(prescribe_chemotherapy) 
│  ├─ WITHOUT intending nausea (sub-world selection)
│  └─ WITH explicit management plan for side-effects
└─ Must inform about side-effects (belief obligation)
   └─ But not commit to wanting them (goal independence)

Critical Insight: The system intends the specific branch where chemotherapy achieves tumor shrinkage with minimal nausea (through sub-world selection), not the entire logical closure including all side-effects.

Alternative Strategies Considered:

• Lower-dose regimen (BELIEVE(lower_dose → less_shrinkage AND less_nausea))
• Combination therapy (BELIEVE(chemo+anti_nausea → shrinkage AND reduced_nausea))
• Sequential treatment (INTEND(try_alternatives_first))

Example 3: Commitment Revision Under Resource Constraints

Scenario: Autonomous research assistant managing multiple paper deadlines with computation budget.

Initial Commitments:

• INTEND(finish_paper_A_by_deadline) [requires 100 GPU hours]
• INTEND(finish_paper_B_by_deadline) [requires 80 GPU hours]
• BELIEVE(available_GPU_hours = 120)

Crisis Point: Halfway through, budget reduced to 60 hours total.

Strategy Comparison:

• Blind Commitment (AI₁): Maintain both intentions despite impossibility
• Single-minded (AI₂): Drop one intention, focus resources
• Open-minded (AI₃): Re-evaluate which paper is still desired more

Expert Decision Process:

1. Impossibility Recognition: BELIEVE(¬(finish_A AND finish_B))
2. Goal Prioritization: Reassess GOAL rankings based on impact
3. Resource Allocation: Choose INTEND(finish_A) OR INTEND(finish_B), not both
4. Fallback Planning: Develop INTEND(submit_preliminary_results) for dropped paper

Trade-off Analysis: Single-minded commitment chosen because collaboration context requires reliability (better to deliver one complete paper than two incomplete ones).

QUALITY GATES

• [ ] Mental state consistency verified: No INTEND(p) without GOAL(p), no GOAL(p) without BELIEVE(optional(p))
• [ ] Sub-world constraints properly applied: Intention-worlds are geometric subsets of goal-worlds, which are subsets of belief-worlds
• [ ] Commitment strategy axioms explicitly chosen and justified for environment type
• [ ] Side-effects handling verified: No unwanted logical closure from BELIEVE(p→q) + INTEND(p) to INTEND(q)
• [ ] Action vs. achievement distinction maintained: does(e) vs. succeeds(e) commitments clearly separated
• [ ] Temporal persistence rules defined: Conditions for maintaining/dropping intentions over time specified
• [ ] Branch vs. world semantics correct: optional/inevitable within worlds, BEL/GOAL/INTEND across worlds
• [ ] Volitional commitment axiom (AI₄) properly implemented: INTEND(does(e)) guarantees execution
• [ ] Goal decomposition strategy validates: Complex intentions broken into achievable sub-intentions
• [ ] Multi-agent coordination model specified: Own intentions vs. beliefs about others' intentions distinguished

NOT-FOR BOUNDARIES

Do NOT use BDI for:

• Pure reactive systems → Use subsumption architecture or behavior trees instead
• Learning-based agents → Use reinforcement learning or neural architectures; BDI assumes fixed logical rules
• Real-time hard constraints → BDI deliberation may be too slow; use real-time scheduling algorithms
• Resource optimization problems → Use operations research methods; BDI doesn't model resource consumption
• Pattern recognition tasks → Use machine learning; BDI is for deliberative reasoning about commitments
• Emotional or social reasoning → Use affective computing models; BDI covers only belief-goal-intention attitudes

Delegate to other skills:

• For planning algorithms: Use classical STRIPS/PDDL planners
• For multi-agent negotiation: Use game theory or auction mechanisms
• For uncertainty quantification: Use Bayesian networks or Dempster-Shafer theory
• For temporal reasoning: Use temporal logic or constraint satisfaction
• For learning from experience: Use reinforcement learning or case-based reasoning

Integration boundaries:

• BDI provides the semantic foundation for deliberative commitment
• Build procedural layers (plan libraries, execution monitoring) on top
• Interface with reactive layers below for real-time response
• Combine with learning systems that update beliefs and goals over time