7
$\begingroup$

If someone wants to develop a basic AI with some code modules,Let us say the AI just has to provide an action when stimulated in a certain situation based on its previous understanding of situations.

I can think of at least 3 of such components:

  • Real-time Understanding/Learning: Using Deep Learning/ConvNets, Supervised/Unsupervised.
  • Logical Decision-Making: Calculating the results of various decisions when applied on current situation based on previous understanding and choosing the most appropriate one logically.
  • Action/Reaction: Acting precisely in the new situation according to the decision-made.

Any ideas?

$\endgroup$
  • 3
    $\begingroup$ Very broad question. Depends how you would define 'basic', how you would restrict its actions, what are the parameters of a 'certain situation', not to mention its 'understanding'. AI's typically have narrowly focused designs for solving specific problems selecting from a diverse set of solution techniques. Consider rephrasing (or abandoning this question entirely) toward a question about a specific problem you want to solve and which types of solutions you're considering applying. $\endgroup$ – dynrepsys Nov 22 '16 at 15:24
  • $\begingroup$ One of the important component is Address space or result space, where the result lies in. $\endgroup$ – Mukul Varshney Nov 24 '16 at 3:14
4
$\begingroup$

Semantics Matters

The answer depends on the definition intelligence being used. If you define intelligence as the ability to adapt, a number of things could be considered intelligent that don't normally fit under the classic AI umbrella.

  • Nonlinear least-squares Marquardt-Levenberg curve fitting algorithm with a substantial but finite set of models, parallel trials of multiple models, outcomes analysis, and smart decisioning
  • Check reader software recently deployed in bank branches and offices
  • The combination of medical providers, patients, and carriers and the modification of treatment through financial instrumentation to improve outcomes

If you define intelligence as mimicking the full array of human mental abilities, no such system is yet available to the public — nothing even close. Creative adaptation across a wide array of arbitrary domains, few of which have been previously experienced in great detail or studied in depth has been an eventual objective of AI research. Approaches have been offered, but none have yet born public fruit. If such a system exists in secrecy, someone would have to violate their nondisclosure agreement or security clearance to tell us about it here.

The definition of intelligence is central to answering. For instance, some reasonable definitions of intelligence might lead an unbiased judge to rate ants above humans. Ants had been building in hexagons for millennia before humans blundered into the habit of building in rectangles. Rectangles require over 70% more building material to build vertical structure per square foot of floor space than a packed hexagonal structure.

Basic Artificial Intelligence System Requirements

Guessing that by, "Basic AI," you mean some naive machine learning, there are a few basic components. (The term Naive in this context means that the AI does not understand the domain or the meanings of symbols or signals it is processing in the way a human who had studied or worked within the field might understand them.)

  • SENSING — The machine (computer) must receive information, generally as a time series. In human beings, these are the senses. In a mail sorter, it may be a camera. This is beyond just the concept of input in information technology. It is more analogous to an input signal in a PID controller. In an automated high speed trading machine, this would be the high speed version of a ticker tape.

  • CONTROL — The machine must manipulate externals in a way that exploits the received information. In a basketball player, this is the motor coordination, facial affect, verbal signals to teammates, and perhaps some verbal bait for opposing players. In a mail sorter, this would be the motor control of mail direction.

  • MEMORY — The machine must have storage to audit input time series (perhaps from some transducer in the real world or some data store upon which some intelligent analysis or transformation must be done). In more advanced systems, the machine may wish to analyze its own performance and make adjustments that converge on some optimal metric value (perhaps a historical maxima or minima) or some range of acceptability.

  • FEEDBACK — The machine must interpret some feedback signalling or use a predetermined scoring mechanism. Learning cannot occur in an information vacuum, so some definition of better or worse must be established. The feedback may be entangled within the SENSING channels or may arrive through a completely separate channel. In biological systems, these are often pre-wired as threat detection, pain, and pleasure. The cerebral cortex uses concepts of goals and progress. In some ways, child rearing and social strata exists to teach the boundaries of what constitutes an acceptable goal and acceptable methods for making progress toward it.

  • MODELLING — Whether implicitly or explicitly, some model must be developed and exploited. Some would say that the existence of a model upon which the predictive capabilities of it can be applied to decision making to achieve some goal or weighted collection of goals is intelligence. Others would say that the development of the model is intelligence and the use of it is merely control mechanics.

Approaches to Simulating Human Thinking

Cognition is not the only form of model making, but the creation of cognitions and their application to decision making. The concept of intelligence may have been furthered along a realistic path by Roger Schank, who proposed that the storage and indexing of stories was a primary characteristics of what humans recognize as intelligent conversation.

Minsky and others took a direction that was more connected with logical inference work that began with logic formalization (originally George Boole) and Church's lambda calculus.

Some Common Directions in Design

The genetic algorithm influenced convergent technology and neuro-biology influenced neural net development. Pattern matching is another limb off the larger set of technical approaches under the umbrella of classic Artificial Intelligence.

These are naive systems. Like a neuron, the components have no idea of the meaning of what they are processing. They are naive components. An intelligent observer could not ascertain the real time meanings of signals and symbols between these naive components without extensive, perhaps life-long research.

Naive Bayesian methods are probabilistic in nature. They exploit Bayes' Theorem, and have been found to produce excellent results in certain important domains. Some studies have shown that naivety is actually a learning accelerant, which is interesting from an AI theory point of view.

Then there is fuzzy (weighted) logic, which was an attempt to merge neural nets with production (rule based) systems. Attempts to use this technology in transportation routing and scheduling has met much success.

There are as many devices and architectures that attempt to effectively integrate or interconnect these various approaches as there are AI projects.

Modelling Environment and Goal Conditions

All of these systems, in some explicit or implicit way, model the external environment and the desired result of system behavior and attempt to converge (in real time) on that result. Some sense-control functionality, which may change and adapt to the external environment is employed in the CONTROL component(s).

This is just another way that the systems have an adaptive behavior. Without necessarily knowing why, the system will manipulate what it can and continue to monitor the state of the environment to continually reach for the system's acceptance criteria.

The Basic AI system must to more than learn. It must also judge its own functionality and therefore must have a layer of feedback and control that simulates the perception of optimality. This higher level control must be integrated into the system at its inception for it to behave intelligently in the sense you probably mean.

Understanding Limitation and System Complexity

The more sophisticated and adaptive the modelling becomes, the more the cognitions, rules, stories, time series coefficients, weights, or whatever forms knowledge (not information) is stored, the more one can say that there is some form of understanding or comprehension.

One conjecture is that it is the recursion in layers of these capabilities that permits certain types of comprehension and awareness. Other conjectures focus more on alertness and attention as keys to higher intelligence.

But these much more mature capabilities are beyond mere adaptation based on past knowledge or information and are therefore beyond what you probably meant by Basic AI.

$\endgroup$
4
$\begingroup$

Some good places to start would be cognitive architectures and as mentioned in another answer intelligent agents. The question is broad but you definitely want to look into planning & decision making. You might also want to check out the L5 and L6 layers of Hierarchical Temporal Memory (As in Nupic) as it relates to feedback, behavior and attention.

If I were you I'd aim for more cognitive solutions (I realize that term is a bit ambiguous itself when we talk about machines). There's also new AI initiative going on involving probabilistic programming. See Probabilistic Models of Cognition made by Goodman (Stanford University) and Tenenbaum (MIT) or Anglican made by Wood (University of Oxford) et al.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.