« Machine learning vide… | Home | Follow me on Twitter »

What artificial intelligence can achieve

26 01 09 - 13:13 The time when artificial intelligence will make robots more intelligent than we are, has not arrived yet. Artificial intelligence is however more than a dream for illuminated scientists; it is a very active and broad research field from which many useful tools for solving problems have arisen.
The applications where artificial intelligence can help are broadly divided into three categories, that are detailed hereafter...

1 - Technologies that enable robots/computers with human-like capabilities.

The way computers and humans interact has changed dramatically since the 80's -- from punched cards* to voice control. Computers are now able to 'hear' and 'see': they can recognize someone from its voice or its face, read printed, or even hand-written texts, write down spoken text, etc. These activites are easy for humans, they are however incredibly difficult for computers. Computers indeed need to be explained precisely what to do and how to do it. They can only follow precise recipes (algorithms). The problem is that even though it is easy for us to recognize people by looking at their faces, we are unable to explain how our brains do the job. It is therefore impossible for programmer to develop software for face recognition directly. What they do is to develop software to teach the computer how to recognize faces.

keywords: computer vision, optical character recognition, speech recognition, natural language processing, face recognition, speaker recognition,

 

2 - Methods for solving hard problems.

Problems can be classified according to how hard (difficult) they are. For instance, finding how much wood is necessary to build a square box of given dimension is an easy problem. Sorting words by alphabetical order is a bit more difficult, but it still is tractable. By contrast, beating an opponent at the game of chess is a much more difficult task. Another very difficult, well known, problem, is the travelling salesman problem; given a set of cities to visit, in which order should they be visited to minimize the total number of kilometers travelled. For a human, solving that sort of problem requires skills and experience. Some are good at palying chess, others are not. The main issue with this sort of problems is the (very) large number of possible solutions and the fact that the optimal solution can only be reached by testing all possibilities. The human mind can be trained to 'prune' solutions, that is refuse to consider entire classes of solutions that we know do not contain the optimal solution. That is what chess players do ; they know some moves will inevitably lead to material loss. Artificial intelligence helps finding the optimal solution or a good approximate solution of such problems using both large computing power and 'tricks' or heuristics, or even randomisation, to accelerate search.

keywords: NP-complete problems, TSP,  heuristic search, planning and scheduling, genetic algorithms, Branch&Bound.

 

3 - Tools for extracting knowledge from facts

Facts can appear under the form of figures (data), or sentences (predicates). From those facts, we often want to extract knowledge, that is general truths that explain the facts. For instance, when we observe that lightning is always followed by thunder, we 'infer' that when we see a lightning, we know that we will soon hear thunder. We can infer how far from us the lightning hit the ground by counting the number of seconds between both events. Computers can be used to infer general equations (models) from a set of data. Once these equations are established, they can subsequently be used to provide predictions. For instance, data can be collected about the number of phone calls recieved at an emergency phone center, and a model can be constructed from those data. The model is then subsequently used to predict how many phone calls will arrive in a few hours, and this prediction is taken into account when setting up empoyees' schedules for the day. The main issue in building models is to know how broadly the model applies. We know that, during special days (national days, etc.) they will probabily be more calls than on regular days. The computer does not; unless we explicitely incorporate that information into the model.

keywords: machine learning, automated reasoning, artificial neural networks

 

Conclusion

In this article, we illustrate the fact that artificial intelligence, albeit being a domaine of hype and fantasy, actually produces tools, methods, and technologies to solve real problems. Artificial intelligence is actually a very broad reserach field, borrowing concepts from mathematics, computer science, psychology, information theory, applied mathematics, and many others.