Abram, I'm giving a talk in a couple weeks at CISIS-12 entitled "Patterns for Cognitive Systems."In the paper I discuss a number of patterns and their functional elements. Attention is anelement of the Motivation pattern. So, if one is implementing the "Motivation" pattern, then you need an attention component to reprioritize your goals (because, by definition, that component is part of the pattern.) Circular argument, but it suffices. The good news is that if you're not implementing motivation the you can exclude it from your architecture. The bottom line becomes, what is the architecture of your cognitive system? In this specific architecture, PAM-P2, I have an attention component, so I'm tryiing to discern useful inputs to the components beyond what is prescribed by the pattern. Is PAM-P2 the best architecture, by no means. It's just one architecture which implementsspecific patterns. For me the question of GraphPlan versus SHOP2 versus CBP is just one of functionality for a specific component within a pattern, namely, the planner. In fact, for that architectural element, the planner, one should be able to seamlessly replace the implementation. This means that at some point I can use a Graphplan planner, or swap it out for a case-based planner, or for an HTN planner, if the interfaces are consistently defined.
Now back to the question at hand... ~PM. Date: Tue, 12 Jun 2012 17:50:24 -0700 Subject: Re: [agi] Attention From: [email protected] To: [email protected] PM, Allow me to ask a question... Should there be such a component? This is my own view, not necessarily strongly supported by the field, but... the way I see it, the literature on classical planning is a graveyard of algorithms based on goal-subgoal hierarchies. Researchers try very hard to get this sort of an approach to work (and succeed), because that is the way we intuitively think we think. However, that doesn't mean it is the best approach. A more bottom-up & global approach may work better. Hierarchical planners were soundly defeated by Graphplan in 1999, an algorithm which treated planning as a constraint satisfaction problem (using a standard solver for those problems) rather than attempting to solve in a pseudo-human way. Hierarchical planners came back, using graphplan as a sub-component, but I see that as more a result of the sticktoitiveness of the hierarchical planning community than a fundamental result. That's in the domain of classical planning. Outside that domain, things like RL and monte-carlo planning are used; subgoal hierarchies do not exist... This could well be for lack of sophistication. Maybe the current 'flat' techniques will be replaced by hierarchical planning in time, even in very messy domains. (I know that hierarchical RL has been implemented more than once, with different approaches...) Certainly I think an advanced AGI system will need the capability to think about plans on multiple levels of abstraction. However, we don't know "where" in the mental architecture this occurs. What I'm saying (or, what I think I'm saying) is that it doesn't seem right to have it at the lowest level. Subgoal prioritization may be a rather advanced form of reasoning which emerges as a result of a lot of lower-level stuff. Of course, something to direct reasoning at the low level is needed! However, it might be a very "different" algorithm from what you expect... Really, I'm just trying to encourage some creative thinking here. :) --Abram On Mon, Jun 11, 2012 at 5:31 PM, Piaget Modeler <[email protected]> wrote: Abram, you've characterized it properly. In my vernacular subgoals = goals. I would say that the job of this particular attention module is to reprioritize the open goal set,given all available information. So the question for me is what should all available information consist of? Some candidates are: (1) The current context, for sure, (2) alerts, (3) expectation failures and mismatches, (4) past prioritizations, (5) past episodes. Anything else? Your thoughts? ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
