Subsumption (Brooks)

Technique 1: Brooks’s three theses

Problem it solves: a short-answer classic. Intelligence

1. can be generated without explicit representations of the symbolic-AI kind;
2. can be generated without explicit abstract reasoning of that kind;
3. is an emergent property of certain complex systems.

Two ideas drive it: situatedness and embodiment (real intelligence is in the world) and intelligence and emergence (it comes from interaction with the environment; it’s “in the eye of the beholder”).

Technique 2: behaviours and inhibition

Problem it solves: how the agent is built.

A subsumption architecture is a hierarchy of task-accomplishing behaviours, each a simple “if situation, then action” rule. Behaviours compete; lower layers are more primitive (avoid obstacles) and take precedence.

Formally (slide 30, book p. 87):

• \(see\) maps states to percepts, as in week 1;
• a behaviour is \((c, a)\) with \(c \subseteq Per\), \(a \in Ac\); it fires in state \(s\) iff \(see(s) \in c\);
• the agent’s behaviours \(R\) come with an inhibition relation \(\prec\), a strict total order. \(b_1 \prec b_2\) means \(b_1\) inhibits \(b_2\).

Technique 3: the action function

Problem it solves: picks the action, very fast.

function action(p : Per) : Ac
    var fired : ℘(R)
begin
    fired := {(c, a) | (c, a) ∈ R and p ∈ c}
    for each (c, a) ∈ fired do
        if not (∃(c', a') ∈ fired such that (c', a') ≺ (c, a)) then
            return a
    return null
end

In words: find every behaviour that fires, and return the action of the one no other firing behaviour inhibits. If none fire, do nothing.

Cost. As written, at most \(n^2\) comparisons for \(n\) behaviours. Since \(\prec\) is a total order you can do better: keep the behaviours sorted and return the first that fires, at most \(n\) checks. Wired into hardware it’s constant time. Compare theorem proving, which can be undecidable.

Technique 4: the Mars explorer

Problem it solves: shows cooperation emerging with no communication at all (Steels; slide 31, book pp. 87 to 90).

Robots collect rock samples that lie in clusters at unknown places. A beacon at the base creates a gradient. Behaviours, highest priority first:

1. If you detect an obstacle, change direction.
2. If carrying samples and at the base, drop them.
3. If carrying samples and not at the base, drop 2 crumbs and travel up the gradient.
4. If you detect a sample, pick it up.
5. If you sense crumbs, pick up 1 crumb and travel down the gradient.
6. If true, move randomly.

Why the crumbs work. A robot carrying a sample home lays a trail. Others who find it follow it away from base, towards the cluster, picking crumbs up as they go. A trail to an exhausted cluster fades; a productive one is refreshed by every robot coming home along it. The robots never talk: the environment carries the information. That’s implicit cooperation (stigmergy), and an early example of the emergence studied in Part III. In the week 3 notebook’s simulation, crumbs raised deliveries from 25.3 to 35.8 samples on average.

Technique 5: the limitations

Problem it solves: the other half of any “evaluate subsumption” answer (slide 32, book 5.1.5).

• Local information only: decisions use the current percept, so no long-term or non-local view.
• Why does it work? Behaviour “emerges”, so the link between rules, world and overall behaviour isn’t understood. You lose BDI’s knowledge-level transparency.
• Hard to debug when it doesn’t work.
• No design method: trial and error, and very hard beyond a handful of layers (the book says fewer than about ten).
• Hard to explain to humans: there’s no symbolic state to report.

Discussion (slide 33): “black-box but effective” today. The same critique fits deep learning and LLM agents, which dominate anyway. A fair position: transparency was traded for capability because it paid off, but the price shows in safety-critical settings, which is why hybrid designs put a transparent layer around an opaque one.

What to read in the book

Practice prompt

I'm studying CS5940 Intelligent Agents (Wooldridge, An Introduction to MultiAgent Systems, 2nd ed., 5.1; Brooks 1991). I want practice on THE SUBSUMPTION ARCHITECTURE:
- Brooks's three theses, situatedness/embodiment and emergence
- behaviours (c, a), when a behaviour fires, the inhibition relation and what b1 < b2 means
- applying the action function: given a prioritised list of behaviours and a percept, which fire and which action is chosen
- the cost of action selection (n^2 as written, n if sorted, constant in hardware) vs theorem proving
- the Mars explorer: which behaviour wins in a given situation, and why crumb trails give implicit cooperation (and fade when a cluster is used up)
- the limitations of purely reactive agents
- whether "black-box but effective" is acceptable today

Please:
1. Use plain, everyday language and explain any notation in one short sentence.
2. Give me ONE question at a time, easy first, then harder.
3. Wait for my answer, then mark it and show a model answer.
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.