4
$\begingroup$

I've read on wiki that already in 2017 there were over 40 institutions researching AGI, and I wonder what type of algorithms are being studied and developed in this field.

For example, for comparison with narrow AI, where models/techniques, such as ANNs, CNNs, SVMs, DT/RT, evolutionary algorithms, or reinforcement learning are used, how would AGI models differ? Do they also use these models but in some specialised way or maybe these algorithms are completely new and different from these currently used in narrow AI?

$\endgroup$
0

2 Answers 2

1
$\begingroup$

Current AGI approaches are very heterogenous and therefore there are no dominant algorithms. Nevertheless, I would suggest you to have a look at Knowledge Graphs and the related algorithms to build and query the graphs, for instance here.

$\endgroup$
1
$\begingroup$

No one has ever invented a practical AGI. However, there are different approaches to the creation of AGI:

  • universalist (e.g. AIXI)
  • symbolic
  • sub-symbolic (e.g. neural networks)
  • hybrid

See also this answer.

So far, the only approach we can say that has really created a (theoretical) AGI is the universalist. The main universalist approach is known as AIXI, a theoretical framework for AGI proposed by Marcus Hutter at the beginning of the 2000s, but it is incomputable, although there are at least 2 computable approximations: AIXItl and MC-AIXI (FAC-CTW), the latter is also practical for simple problems, like tic-tac-toe. Another universalist solution is the Godel machine by Schmidhuber. Here is an answer that explores the importance of AIXI to the AI field.

So, if you really want to learn about and do research on AGI, you could start with AIXI or Godel machines. However, to understand them, you need to have some basic understanding of RL, theory of computation, probability theory, and other concepts. So, you could also start with reviewing these concepts, before getting into AIXI.

Having said that, people that are doing research on RL, neural networks, evolutionary algorithms, cognitive architectures, or other AI topics might have the ultimate goal of developing an AGI. Some people might think that some kind of reinforcement learning (see this) and neural networks (see this) to represent the learned information is necessary for developing an AGI, because humans and other animals also learn by reinforcement and have neural circuits. Other people think that we need some kind of cognitive architecture and some means of doing logical thinking. Other people think that we need to reverse engineer the neocortex (see Numenta's work) or that we need neuromorphic chips.

So, although our current approaches to RL and our current models have only been used to solve narrow problems, I think it is possible that these or advanced/modified versions of them can be used to create a practical AGI.

The other answer mentions knowledge graphs, but KGs have only been used to solve narrow problems too, e.g. they are used by Google to display specific information about important people when you search for them on their engine. They are also probably used in personal assistants, like Alexa, Siri and Google Assistant.

In short, if you're interested in AGI algorithms and models, look into AIXI, Godel machines, Numenta's work, cognitive architectures (like SOAR or OpenCog), in addition to the more traditional ML approaches, like RL. I also recommend that you read the paper Artificial General Intelligence: Concept, State of the Art, and Future Prospects.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .