Intention reconsideration
How often should an agent stop and ask "is this still the right goal?" Never, and it chases goals that no longer matter. Always, and it never gets anything done. The answer depends on how fast the world changes.
Kind: deliberative agent (BDI meta-level control).
Week: 3 (slides 12 to 14).
Book: Wooldridge 4.3, pp. 77 to 79 (Table 4.2, Kinny and Georgeff).
Techniques on this page: cycle (5) with reconsider, when reconsider is optimal, bold vs cautious agents, the maths of dynamism, ending the meta-level regress.
The problem it solves
Cycle (4) only rethinks its intentions between plans. If a better opportunity turns up mid-plan, or the goal quietly stops making sense, it won’t notice.
The two obvious fixes both fail (book p. 77). Never reconsider and the agent keeps pursuing intentions that can’t succeed or no longer matter. Reconsider after every action and, since deliberation is expensive, it spends its time thinking instead of acting.
The solution is a cheap test that decides whether to deliberate. How eager that test should be is the Kinny and Georgeff experiment: one of the few real experimental results in the module, and good exam material.
Words you’ll hear
| Word or symbol | Plain meaning |
|---|---|
| meta-level control | a decision about whether to make a decision |
| \(reconsider(I, B)\) | a cheap yes/no test: should I deliberate now? |
| boldness \(k\) | the most plan steps the agent runs before reconsidering |
| bold agent | never reconsiders until the plan is finished (\(k = \infty\)) |
| cautious agent | reconsiders after every action (\(k = 1\)) |
| dynamism \(\gamma\) (gamma) | how fast the world changes: changes per pass of the control loop |
| effectiveness | intentions achieved ÷ intentions the agent had (1 = everything) |
| Poisson process | events that happen at random at a steady average rate |
Technique 1: cycle (5)
Problem it solves: adds controlled reconsideration of ends, separate from repairing means.
9. while not (empty(π) or succeeded(I, B) or impossible(I, B)) do
10. α := head(π); 11. execute(α); 12. π := tail(π);
13. get next percept ρ; 14. B := brf(B, ρ);
15. if reconsider(I, B) then /* commitment to ends, controlled */
16. D := options(B, I);
17. I := filter(B, D, I);
18. end-if
19. if not sound(π, I, B) then /* commitment to means */
20. π := plan(B, I);
21. end-if
22. end-whileWhy this order? Lines 15 to 18 may change \(I\) but don’t replan. If \(I\) changed, the old plan no longer achieves it, so \(sound(\pi, I, B)\) fails and lines 19 to 21 replan. Rethinking the goal and repairing the plan are kept separate.
Technique 2: when is reconsider optimal?
Problem it solves: a precise definition of “a good reconsider function” (book Table 4.2).
Assume deliberation always picks the best intentions and time always costs something.
| Situation | Chose to deliberate? | Would deliberating change intentions? | Optimal? |
|---|---|---|---|
| 1 | no | no | yes |
| 2 | no | yes | no: missed a change |
| 3 | yes | no | no: wasted effort |
| 4 | yes | yes | yes |
So optimal ⟺ it deliberates exactly when deliberating would change the intentions.
The hidden assumption: \(reconsider\) must be much cheaper than deliberation. Otherwise it could just run deliberation inside itself as an oracle, which would cost as much as always deliberating.
Technique 3: bold vs cautious agents
Problem it solves: which reconsideration policy to choose for a given environment (Kinny and Georgeff, 1991).
- Low \(\gamma\) (calm world): bold agents win. Cautious agents waste time rethinking commitments that are still fine.
- High \(\gamma\) (fast-changing world): cautious agents win. They spot doomed intentions early and grab new opportunities.
The book’s bottom line: different environments need different strategies. And the cost of deliberation matters too: make it expensive enough and cautious loses everywhere.
Technique 4: the maths of dynamism
Problem it solves: puts a number on why a bold agent’s plans go stale.
Model world changes as random events at rate \(\gamma\) per step. A bold agent commits to an \(n\)-step plan and doesn’t look again. The plan survives only if nothing changes for \(n\) steps:
\[P(\text{no change in } n \text{ steps}) = e^{-\gamma n}.\]
| \(\gamma\) | \(n = 5\) | \(n = 10\) | \(n = 20\) |
|---|---|---|---|
| 0.02 | 0.905 | 0.819 | 0.670 |
| 0.1 | 0.607 | 0.368 | 0.135 |
| 0.4 | 0.135 | 0.018 | 0.0003 |
Worked example. \(\gamma = 0.05\), 8-step plans: \(e^{-0.05 \times 8} = e^{-0.4} \approx 0.670\). A third of the bold agent’s plans are stale before they finish.
A cautious agent instead pays a deliberation cost \(d\) on every step: \(n \cdot d\) per plan rather than \(d\). With \(d\) = a quarter of a step and \(n = 10\), that’s 2.5 extra steps per plan. Expensive in a calm world; cheap compared with walking to a target that’s already gone in a busy one.
Technique 5: ending the regress
Problem it solves: the slide 14 question: deciding whether to reconsider is itself a decision, so do we need a meta-meta-level, forever?
A good answer: the regress stops at something fixed and cheap. \(reconsider\) isn’t another deliberator; it’s a hard-wired, constant-time test, like “reconsider only if a belief my intention depends on has changed”, “every \(k\) steps”, or “when a surprising percept arrives”. The designer picks it once, offline, from what they know about the environment’s dynamism (exactly what Kinny and Georgeff measured). It’s the same reason reactive rules exist.
With an LLM as the planner (the week 3 notebook): a cautious agent calls the LLM after every step (6 calls, about 3× the tokens); a bold one is cheap but walks into a closed door; a single-minded agent with a cheap \(sound()\) check gets both: no wasted actions and only 2 calls. That’s Table 4.2’s hidden assumption in practice.
Notation. Slide 13 writes dynamism as \(\gamma\). The book doesn’t use the symbol; it says “rate of world change”.
What to read in the book
| Section | Read? | Why |
|---|---|---|
| 4.3, pp. 77 to 79 | Read | Reconsideration, Table 4.2 (p. 78), Kinny and Georgeff’s results (pp. 78 to 79) |
| Ch. 4 notes | Skip | References only |
Page numbers are the book’s; in the scanned PDF add 21.
Practice prompt
I'm studying CS5940 Intelligent Agents (Wooldridge, An Introduction to MultiAgent Systems, 2nd ed., 4.3 pp. 77-79). I want practice on INTENTION RECONSIDERATION:
- cycle (5): where reconsider, options, filter, sound and plan sit, and why reconsidering ends comes before repairing means
- the four situations of Table 4.2 and when reconsider is optimal; why reconsider must be much cheaper than deliberation
- bold vs cautious agents, boldness k, dynamism gamma, effectiveness, and Kinny and Georgeff's results
- the probability a bold agent's n-step plan survives, e^(-gamma n), and comparing it with a cautious agent's deliberation cost n*d
- choosing a reconsideration strategy for a described environment
- why the meta-level regress stops at a cheap, fixed test
- the cost of cautious reconsideration when the planner is an LLM (tokens per call)
Please:
1. Use plain, everyday language and explain any symbol in one short sentence.
2. Give me ONE question at a time, easy first, then harder, with friendly numbers.
3. Wait for my answer, then mark it and show the full working.
4. If I get it wrong, explain the mistake and give me a similar question.
5. Every 5 questions, summarise my strengths and what to practise.