AACA Textbook
From AGIRI.org
Contents |
Adaptive Agoric Computer Architecture Textbook
Preface
I'm attempting to do the same as Instead of an AGI Textbook but I would like to put in more editorialising about why each subject is important for my particular approach, as learning is easier when you know why you are doing it.
I have come to the conclusion that building an intelligence will have to be done in stages. That current computers architectures aren't quite correct for building an intelligence upon, but keeping the same flexibility and ease of programming as them is important. All the reading will be on the first stage, as that is what has to be built first. What reading is needed for the second stage or further stages will depend upon how the first stage goes and is not my area of expertise. Feel free to merge the topics I create here into the main wiki article.
-- Will Pearson
Table of Contents
Inspirational Neuroscience
I am trying to make the computer system have the same low level features that makes brain able to support human like intelligence (as well as the other possible brains)
Very low level changes in the brains wiring suggests the level at which the intelligent system has to be able to change.
This shows that the brain as well as having high level thought and reasoning, has very low level control of energy flows into it.
Shows the brain is not rational but is arranged around needs and wants
Computer Science and Mathematics Background
Knowing that pure reason cannot get you everywhere is important.
To show that the space of possible programs in a computer is not equal to the space of mapping between the inputs and outputs of that computer
Machine learning and reinforcement Learning
Most Search Spaces in machine learning are hypothesis spaces, that is hypothesis about the outside world. This requires some (generally fixed) way of translating the external stimuli to internal concepts. In the Architecture I am interested in, the space to be searched is the state space of the computer architecture. Or to put it another way it tries to find the way system should "be", rather than trying to find about the outside world. Why?
Because the No Free Lunch Theorems say there is no particularly good way to infer which function you are dealing with out of all possible functions from information from that trials from that function. As we don't know much about the world, we shouldn't fix any bias about how the world is, in a learning system. Also inference is not enough for quick learning, you have to have quick assimilation/instruction accepting.
Reinforcement Learning Methods
Reinforcement Learning can be used because it needn't just refer to how well the system understands the world, it can refer in general to how "good" the system is at this point in time.
Evolution & Economics
So how to get the system to have a sense of good without restricting the type of programs and interactions between them?
Hayek and Economic Credit Assignment
Security
While the Architecture shouldn't restrict what the programs can do, the programs themselves are free to restrict how they interact. The problem of how to allow this form of restriction is generally placed under the rubrick of security. The cleanest form of security I have come across is object capability based systems.
