3 ms·
Subsumption seemed like a cool idea when it was introduced, but it turns out that it does not actually work. http://www.cs.uml.edu/~holly/91.549/readings/tla.p
by lisper 10y ago
Subsumption seemed like a cool idea when it was introduced, but it turns out that it does not actually work.
http://www.cs.uml.edu/~holly/91.549/readings/tla.pdf http://www.cs.uml.edu/~holly/91.549/readings/tla.pdf
- Animats 10y agoBack then, I said that we needed more structure than Brooks, but less structure than Latoumbe (a professor at Stanford into efficient exhaustive search planning). Before Brooks, there was a long tradition of detailed planning in AI. Brooks totally rejected planning. This is an extreme position, but the planning and re-planning people were having trouble coping with the real world. Brooks' Roomba is an expression of that model. The Roomba has a few simple behaviors - go in straight line, spiral, follow wall - and a state machine to switch between them. It has no idea where it is. Most later robot vacuums have actual navigation, because Roombas are notorious for getting themselves into the same traps over and over. Both overplanned and purely reactive systems have trouble coping with the real world, but they fail in different ways.
- lisper 10y ago> Latoumbe I presume you mean Latombe? http://ai.stanford.edu/~latombe/ http://ai.stanford.edu/~latombe/ > This is an extreme position, but the planning and re-planning people were having trouble coping with the real world. Yes. The whole field wasted decades running afoul of Ron's First Law: all extreme positions are wrong. This is why the three-layer architecture described in the link above (reactive behaviors sequenced by a state machine informed by a deliberative planner) eventually became the de facto standard, because it combines both approaches in a way that actually fits the structure of the problem.
- gohrt 10y agoheh, that's an old paper: http://www.flownet.com/ron/eg-rg-faq.html http://www.flownet.com/ron/eg-rg-faq.html
- lisper 10y ago1998 (though it is remarkable how difficult it is to find that date), so very nearly 20 years. But it's a survey; the work described in that paper goes back as far as 1990 or so. (I'm the author, in case you didn't know.)
- infinite8s 10y agoI think the issue with subsumption is that 3layers aren't enough. It would be interesting to combine recent deep neuralnet architectures with the Dr Brooks overall subsumption design (so multiple layers of individual deepnets layered hierarchically)
- lisper 10y agoNo, the problem with subsumption is that it is a poor composition mechanism. Even the name reveals this to be true: a higher layer has to subsume a lower layer, because (according to Brooks) when a higher layer is active the lower layer's outputs are suppressed. As long as the higher layer is active, the lower layer contributes nothing to the final result and hence may as well not exist. In fact, it's really not a "layered" architecture at all, it's just a big cascading series of if-then-else clauses. Subsumption doesn't really do anything other than obscure this fact.
- ChuckMcM 10y agoI think that is overly harsh, the beauty of it for me continues to be that it is default stable. No input and things are stable. That makes it very resilient in low compute environments. That is the same reason I would recommend it for a smart house architecture, when the upper (planning) layer fails (or is broken) you can always get the lower layers to do what you want (lights turn on, or thermostat adjusts) because the local command the top of the 'if-then' tree in your lexicon.
- lisper 10y ago> it is default stable But it isn't. In order for two layers to interact properly you have to specifically engineer them that way. If a higher layer starts or stops subsuming at the wrong time you can get catastrophic failures. Also, smart houses are a very different problem than autonomous mobile robots.
- infinite8s 10y agoAh then I've misunderstood the architecture all these years. I had always thought it was modeled on complex nervous systems.