The List
This list originates from Bruce Maxim, Professor of Engineering, Computer and Information Science at the University of Michigan. In his lecture Spring 1998 notes for CIS 4791, the following list was called,
"Good Problems For Artificial Intelligence."
Decomposable to easier problems
Solution steps can be ignored or undone
Predictable Problem Universe
Good Solutions are obvious
Internally consistent knowledge base (KB)
Requires lots of knowledge or uses knowledge to constrain solutions
Interactive
It has since evolved into this.
Decomposable to smaller or easier problems
Solution steps can be ignored or undone
Predictable problem universe
Good solutions are obvious
Uses internally consistent knowledge base
Requires lots of knowledge or uses knowledge to constrain solutions
Requires periodic interaction between human and computer
What it is
His list was never intended to be a list of AI problem categories as an initial branch point for solution approaches or a, "heuristic technique designed to speed up the process of finding a satisfactory solution."
Maxim never added this list into any of his academic publications, and there are reasons why.
The list is heterogeneous. It contains methods, global characteristics, challenges, and conceptual approaches mixed into one list as if they were like elements. This is not a shortcoming for a list of, "Good problems for AI," but as a formal statement of AI problem characteristics or categories, it lacks the necessary rigor. Maxim certainly did not represent it as a, "7 AI problem characteristics," list.
It is certainly not a, "7 AI problem characteristics," list.
Are There Any Category or Characteristics Lists?
There is no good category list for AI problems because if one created one, it would be easy to think of one of the millions of problems that human brains have solved that don't fit into any of the categories or sit on the boundaries of two or more categories.
It is conceivable to develop a problem characteristics list, and it may be inspired by Maxim's Good Problems for AI list. It is also conceivable to develop an initial approaches list. Then one might draw arrows from the characteristics in the first list to the best prospects for approaches in the second list. That would make for a good article for publication if dealt with comprehensively and rigorously.
An Initial High Level Characteristics to Approaches List
Here is a list of questions that an experienced AI architect may ask to elucidate high level system requirements prior to selecting an approaches.
- Is the task essentially static in that once it operates it is likely to require no significant adjustments? If this is the case, then AI may be most useful in the design, fabrication, and configuration of the system (potentially including the training of its parameters).
- If not, is the task essentially variable in a way that control theory developed in the early 20th century can adapt to the variance? If so, then AI may also be similarly useful in procurement.
- If not, then the system may possess sufficient nonlinear and temporal complexity that intelligence may be required. Then the question becomes whether the phenomenon is controllable at all. If so, then AI techniques must be employed in real time after deployment.
Effective Approach to Architecture
If one frames the design, fabrication, and configuration steps in isolation, the same process can be followed to determine what role AI might play, and this can be done recursively as one decomposes the overall productization of ideas down to things like the design of an A-to-D converter, or the convolution kernel size to use in a particular stage of computer vision.
As with other control system design, with AI, determine your available inputs and your desired output and apply basic engineering concepts. Thinking that engineering discipline has changed because of expert systems or artificial nets is a mistake, at least for now.
Nothing has significantly changed in control system engineering because AI and control system engineering share a common origin. We just have additional components from which we can select and additional theory to employ in design, construction, and quality control.
Rank, Dimensionality, and Topology
Regarding the rank and dimensions of signals, tensors, and messages within an AI systems, Cartesian dimensionality is not always the correct concept to characterize the discrete qualities of internals as we approach simulations of various mental qualities of the human brain. Topology is often the key area of mathematics that most correctly models the kinds of variety we see in human intelligence we wish to develop artificially in systems.
More interestingly, topology may be the key to developing new types of intelligence for which neither computers nor human brains are well equipt.
References
http://groups.umd.umich.edu/cis/course.des/cis479/lectures/htm.zip
