The "architecture" is incremental: no blocks, just levels. All I have is incremental-complexity comparison-evaluation operations, explained in the intro. Everything else is learned, that's what makes it *general*.
From: Piaget Modeler Sent: Wednesday, May 15, 2013 1:15 PM To: AGI Subject: RE: [agi] Goal Selection Perhaps architecture diagram would be a more specific term. What architecture implements your proposed system. Can you draw it? Thanks, ~PM -------------------------------------------------------------------------------- From: [email protected] To: [email protected] Subject: Re: [agi] Goal Selection Date: Wed, 15 May 2013 09:51:48 -0400 Michael, > Do you have a diagram to go along with your explanation? Depends what you're looking for. In humans, every part of the brain affects motivation & attention. For GI, the only useful motive is curiosity (a drive to maximize predictive correspondence), & "discrete components" are learned patterns, rather than built-in "modules". The nearest thing to a diagram that I have is grouping levels of search by incremental order of distance between comparands: - Comparison of adjacent inputs, forming continuous patterns of incremental dimensionality (distance = 1): line segments: 1D, blobs: 2D, objects: 3D, & processes: 4D. This is similar to connected-component analysis. - Cross-comparison across a whole queue of inputs, forming discontinuous patterns per input.(distance->n). These patterns are fuzzy, & their overlap is compensated by selection among inputs to a next level. - Comparison across a hierarchy of short-cuts to higher-level queues, generated by feedback (distance -> nn)...: www.cognitivealgorithm.info. > I'm looking at attention rather than motivation. > For me, attention is the filtering or re-prioritization of goals. Attention is prioritizing areas of search, according to the weights assigned by combined motives (salience). Conscious attention is WTA mechanism, imposed by our brain-to-body bottleneck: http://cognitive-focus.blogspot.com/2012/06/temporal-attention-span-our-dominant.html This is irrelevant for GI because it shouldn't assume any fixed "body". What's relevant is distributed "unconscious" attention, which is a market-like mechanism that allocates cognitive resources to the areas of search, in proportion to their projected contribution to total predictive correspondence of one's model of the environment. > In the PAM-P2 system I have an intuition that a higher level of selection > occurs than is explained than by basic action selection. Right, this is a selection of "cognitive actions": prioritization of internal search, covered above. > Motivation is handled in PAM-P2 through the use of homeostatic variables and > "urges", > deltas between current and target homeostatic variable values. That's an equivalent of my "instincts", - a supervised learning part, irrelevant for GI per se. > My intuition tells me that there should be another, higher level of goal > selection. (Or, perhaps goal filtering, not exactly sure). Something that operates above action selection that takes into account all the possible goals the system could have and ensures that the most important at the current moment are part of the agenda. Again, you're talking about the same thing from different POVs. "Goal" is a positive-value-charged state, or its internalized representation. Motivation is what does this value-charging, thus determines the priority of searching though related internal representations & external sources. From: Piaget Modeler Sent: Tuesday, May 14, 2013 5:56 PM To: AGI Subject: RE: [agi] Goal Selection Hi Boris, Thanks for the references. Do you have a diagram to go along with your explanation? That would be much appreciated. A diagram helps the explanatory cloud to be decomposed into discrete components. I'm looking at attention rather than motivation. Motivation is handled in PAM-P2 through the use of homeostatic variables and "urges", deltas between current and target homeostatic variable values. For me, attention is the filtering or re-prioritization of goals. In the PAM-P2 system I have an intuition that a higher level of selection occurs than is explained than by basic action selection. In PAM-P2 there are two action selectors: the Reactor, which matches existing solutions to sensory stimuli, and the Deliberator which matches existing solutions with active situations and needs (goals). Both action selectors operate in a case based manner, where "solutions" are the cases. Once a solution is selected, it may generate subgoals to assist in attaining the overall solution. My intuition tells me that there should be another, higher level of goal selection. (Or, perhaps goal filtering, not exactly sure). Something that operates above action selection that takes into account all the possible goals the system could have and ensures that the most important at the current moment are part of the agenda. Your thoughts? ~PM -------------------------------------------------------------------------------- From: [email protected] To: [email protected] Subject: Re: [agi] Goal Selection Date: Tue, 14 May 2013 17:15:51 -0400 You're really trying to understand how human motivation works. I already posted this, but in case you missed: Human motivation: developmental perspective. Motivation is all mental mechanisms that drive our behavior, in which I include cognitive behavior: analysis, introspection, & planning for somatic behavior. Values / motives in humans & higher animals can be divided into three broad categories, according to the mechanism that formed or selected them: Evolution selects instincts fit for their own propagation, innate but subsequently modulated by usage, Conditioning value-charges stimuli coincident with previously value-loaded stimuli in time or space, Cognitive curiosity searches / selects for predictive patterns, even if they consist of value-free stimuli. Higher mechanisms accelerate adaptive value acquisition by acting on increasingly mediated responses: from immediate behavioral reactions to longer-term attention, prediction, & planning. Brain areas that implement these value-acquisition mechanisms likely evolved in the same sequence: Instincts, largely physiological & traceable to 4Fs, are encoded mainly in brainstem & hypothalamus. Conditioning is initiated by basal ganglia & limbic system, then extended & generalized by neocortex. Predictive curiosity is an innate driver of neocortex, which is also heavily modulated by lower motives. This scheme is vaguely similar to triune brain model, but in my interpretation these substrates differ mainly in the mechanism by which they acquire values, rather than in resulting & relatively transient motives themselves. These value acquisition mechanisms are innate, but their relative strength varies. Our instincts are pretty basic & similar to those of other mammals. An excellent account of that level of motivation is Jaak Panksepp‘s “Archaeology of Mind: Neuroevolutionary Origins of Human Emotions“. The discussion below is mostly on conditioning & cognition: increasingly adaptive mechanisms which seem to strengthen with our personal growth: http://cognitive-focus.blogspot.com/2012/06/motivation-evolution-of-value.html From: Piaget Modeler Sent: Tuesday, May 14, 2013 4:17 PM To: AGI Subject: RE: [agi] Goal Selection Getting Closer: Top-down versus bottom up attentional control: a failed theoretical dichotomy http://ems.psy.vu.nl/userpages/theeuwes/Trends_2012_Awh.pdf The priority map notion is closer to what I was looking for. I know that priorities fit in somehow. ~PM AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription AGI | Archives | Modify Your Subscription ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
