Q&A: Phase 24.2 — IntentionRecognizer

[web3guru888]

Q&A: Phase 24.2 — IntentionRecognizer 🎯

Q1: How does Bayesian inverse planning differ from simple pattern matching?

A: Simple pattern matching checks if an action sequence matches a known plan template. Bayesian inverse planning treats intention inference as probabilistic inference: P(goal | actions) ∝ P(actions | goal) × P(goal). This means:

• Multiple hypotheses compete simultaneously, weighted by evidence
• Prior probabilities (from context, beliefs) influence inference
• Novel action sequences can support goals even without exact template match
• Confidence naturally degrades when evidence is ambiguous

Q2: What happens when an agent's behavior matches multiple plan templates?

A: All matching templates generate competing hypotheses. The Bayesian framework maintains all of them with calibrated posteriors. After each observation, posteriors update and normalize (sum to 1 across competing hypotheses for the same agent). Over time, the true goal accumulates more evidence and dominates. The top_k parameter in infer_intention controls how many alternatives to surface.

Q3: How does the system handle deceptive agents who deliberately perform misleading actions?

A: The current design detects deception indirectly:

1. If an agent's actions consistently support goal G but then abruptly diverge, the ACTIVE→BLOCKED/ABANDONED transition fires
2. A pattern of goal changes suggests strategic deception
3. PerspectiveTaker (24.3) can simulate "does this agent know we're watching?" for higher-level deception modeling
   The IntentionRecognizer itself is descriptive (what do observed actions suggest?) rather than adversarial. Deception detection is a SocialOrchestrator (24.5) concern.

Q4: How is the plan library populated and maintained?

A: The plan library is populated via register_plan_template() at initialization and can be extended at runtime. Sources include:

• Domain-specific templates (negotiation plans, resource gathering sequences)
• Learned templates from historical behavior analysis
• Templates derived from the agent's own planning (if I would plan this way, others might too)
  Templates with consistently low match rates are candidates for pruning.

Q5: What is the computational cost of maintaining hypotheses for many agents?

A: Per agent: O(H × T) per observation, where H = active hypotheses, T = templates scanned. With pruning (min_confidence threshold), H stays bounded. For N agents, total is O(N × H × T). The bounded deque for action logs (default 1000) prevents unbounded memory growth. The asyncio.Lock is per-agent (not global), so multi-agent updates proceed concurrently.

Q6: How does the GoalStatus state machine work?

A: State transitions:

• HYPOTHETICAL → ACTIVE: confidence exceeds activation threshold (0.3)
• ACTIVE → BLOCKED: expected next action doesn't occur within timeout
• BLOCKED → ACTIVE: agent resumes actions consistent with the goal
• ACTIVE → ACHIEVED: all postconditions of the plan template are satisfied
• ACTIVE/BLOCKED → ABANDONED: confidence drops below pruning threshold
  The distinction between BLOCKED and ABANDONED matters: blocked goals may resume, abandoned goals are pruned.

Q7: How does IntentionRecognizer integrate with BeliefTracker (24.1)?

A: BeliefTracker provides contextual priors for goal inference. If we believe agent A thinks resource X is scarce (from their belief model), goals involving acquiring X get higher priors. The integration:

belief_state = await self._belief_tracker.get_agent_model(agent_id)
goal_priors = compute_priors_from_beliefs(belief_state, plan_library)

This makes intention inference context-sensitive—the same actions can suggest different goals depending on what the agent believes about the world.