Slashdot Mirror

Indeed there are no easy solutions, but there's plenty of mathematical work going on to better handle uncertainty. For example, OpenCog's Probabilistic Logic Networks. From http://wiki.opencog.org/w/Probabilistic_Logic_Networks "PLN is a novel conceptual, mathematical and computational approach to uncertain inference. In order to carry out effective reasoning in real-world circumstances, AI software must robustly handle uncertainty."

I've been noticing a lot of China outreach from projects like OpenCog. Any others?

As humans, we do exactly what physics mandates we do. Unless you're purporting that the human brain uses some sort of hypercomputation or that there's something special (ie outside of our current understanding of physics) about what neurons do, you're not being consistent.

That said, I understand where you're coming from; most AI research is in very narrow domains and has no intention or hopes of solving the problem of achieving human-level intelligence (Watson is an example of narrow AI, as it clearly lacks a genuine understanding of the question or the english language). But the fact remains, that is how the term AI is used.

There's a growing separation between this "Narrow AI" and the kind of AI you seem to be hoping for, Artificial General Intelligence (AGI). There are some AGI projects out there, such as the open source opencog. Since there's no hope of people stopping calling things like computer chess AI, I prefer to use the term AGI whenever referencing "real" AI.

In order to achieve SuperIntelligence an artificial general intelligence (AGI) needs the superfast speed and the massive parallelism of a Supercomputer.

Although the idea of development standards in Artificial General Intelligence (AGI) or AI Standards in general is something of a misnomer for an explosively evolving phenomenon, there are still standards of excellence to be applied in the creating and coding of an AGI. One optional standard is the choice of 64-bit computing platforms as an ideal environment for a machine intelligence requiring random access to a practically unlimited memory space.

Part of the approaching Technological Singularity will be the dislodging of Big Pharma and Big Physics and other traditional supercomputer users from their station as the overlords of High Performance Computing (HPC). AGI will assume its rightful place at the summit of supercomputer usership and ownership. "All your supercomputer will belong to us." The new AGI overlords will not tolerate jonesing among nations for bragging rights to the fastest or biggest Supercomputer on Earth.

One way to boost complexity is to evolve programs.. or neural networks, if you're that way inclined. One way to speed up the evolutionary process is to use probabilistic modeling to produce offspring.. it's must more efficient than just random mutation. See http://www.opencog.org/wiki/MOSES. Eventually, though, you will reach limits to blind search. At that point you need to complement it with logic. See http://www.opencog.org/wiki/PLN. And to focus your search you really need some kind of attention allocation. See http://www.opencog.org/wiki/Attention_allocation. An integrative approach means you can solve real world applications with modest hardware.

One way to boost complexity is to evolve programs.. or neural networks, if you're that way inclined. One way to speed up the evolutionary process is to use probabilistic modeling to produce offspring.. it's must more efficient than just random mutation. See http://www.opencog.org/wiki/MOSES. Eventually, though, you will reach limits to blind search. At that point you need to complement it with logic. See http://www.opencog.org/wiki/PLN. And to focus your search you really need some kind of attention allocation. See http://www.opencog.org/wiki/Attention_allocation. An integrative approach means you can solve real world applications with modest hardware.

One way to boost complexity is to evolve programs.. or neural networks, if you're that way inclined. One way to speed up the evolutionary process is to use probabilistic modeling to produce offspring.. it's must more efficient than just random mutation. See http://www.opencog.org/wiki/MOSES. Eventually, though, you will reach limits to blind search. At that point you need to complement it with logic. See http://www.opencog.org/wiki/PLN. And to focus your search you really need some kind of attention allocation. See http://www.opencog.org/wiki/Attention_allocation. An integrative approach means you can solve real world applications with modest hardware.

http://www.opencog.org/wiki/OpenCogPrime:WikiBook

Some interesting stuff.

http://www.opencog.org/wiki/OpenCogPrime:EssentialSynergies

That's pruning.. To take it out of massive AI geek speak:

    "I need to figure out a way to open this bottle. Out of all the knowledge I have about the world, what should I consider? Should I consider things I know about flowers? Well no. Duh. Should I consider things I know about cats? Well no. Duh. Maybe, I should consider things about *bottles* and gee, I don't know, what else.. hmm.. how about the parts of the bottle, maybe a common part of a bottle that most bottles have: lids. Now, do I know anything about *opening*? Why yes, I do.. I should probably look at all the ways I know to open something and see if any of the things I've opened in the past are like bottles. Maybe I should try to determine what *type* of lid this bottle has on it."

Chess tells you have to conduct a search in a space.. when the space is really really large you need to prune, so how do you do that? By more searching? No.. you do it by looking at what is associated with the objects in your current estimation of the world state and what they imply. You try to infer what motives or other wise *causes* there are for the current state of the world and you use this inferred information to prune. Is there a chess program in existence that does this? No.. not at all. If you read a chess book now and then you'll see all these *concepts* of chess play. All these little tactics like pinning and forks and discovery and pawn development, etc, etc. Where's all the *reasoning* about these concepts in chess programs? Oh, that stuff, that's all in the hard coded board evaluation function.. that's a *given*. This is why chess programs will assign +5 to a board configuration where they gain a pawn, even if winning a pawn really doesn't mean shit right now because you're about to take their queen. If the pruning causes the search not to find the board position where the queen is being threatened, that pawn capturing move is the shit and that's what the next move will be. This is clearly *retarded*. It's nothing like intelligence. People don't randomly consider "what will happen next if I make this move".. they build theories about the other player's strategy and they plan how to foil that strategy whilst developing their own strategy and they make moves that will cause the other player to think they have a strategy that is different to their actual strategy. They set up traps based on their *theory of mind* of the other player and work the other player to get them into the trap. This is why all the AI masters of old thought that chess would be an interesting problem because they imagined that they'd actually get to code up some of this stuff.. instead we got move-space search algorithms with static board evaluation pruning. Yawn.

Not that figuring out that you can represent the position of all the pawns on the board in a single machine word and using bitwise operations you can very quickly calculate the effect of a non-taking non-double move in a single instruction isn't *fun*. Optimizing code is great fun. It's awesome that there's still people in the world who value this stuff. It's just not intelligence.

http://www.opencog.org/wiki/Main_Page
http://www.agiri.org/OpenCog_AGI-08.pdf
http://justingibbs.com/how-to-make-singularity-bearable-in-its-infancy
http://www.innergybv.biz/blog/?p=175
http://ieet.org/index.php/IEET/more/goertzel20080620/#When:22:49:00Z
http://xlaurent.blogspot.com/2008/06/opensim-for-opencog.html

There's a number of GSoC projects for OpenCog currently underway also:

http://code.google.com/soc/2008/siai/about.html

So the first release should be very interesting.

http://cognitrn.psych.indiana.edu/rgoldsto/labware.html
http://ccrg.cs.memphis.edu/projects.html
http://www.opencog.org/wiki/Main_Page
helps to know what you're looking for