Nowadays, not even the best planning algorithms have managed to overcome the human capacity to solve problems. Therefore, specialists are still trying to improve them.
The last attempt has been made by researchers from the Massachusetts Institute of Technology (MIT, USA). They have done it in a surprising way: adding human intuition to the algorithms.
To do this, they have codified the strategies of human planners with high performance, to make them readable for machines. In this way, they have managed to improve the performance of the planning algorithms by 10 to 15%, according to the MIT website .
Problems with different complexities
The people chosen as examples have been individuals who excelled in planning, programming and optimization, in total 36 MIT students. Their strategies were introduced in the form of algorithms introduced in machines so that they solve problems with different degrees of difficulty.
One of the simplest would be, for example, “given a certain number of airports, a certain number of aircraft and a certain number of people at each airport with certain destinations, is it possible to plan flight routes that allow all passengers to arrive to their destinations without any plane flying at the same time? ”
A more complex class of problems are the numerical ones, which consist of adding flexible numerical parameters: “Can you develop a set of flight plans that meet the restrictions of the original problem but also minimize the flight time of the aircraft and the fuel consumption? ”
Finally, the more complex problems – the temporary problems – add temporal limitations to the numerical problems:” Can you minimize flight time and fuel consumption while ensuring that the planes arrive and depart at hours? concrete? ”
]For each problem, the algorithm has half an hour to generate a plan. The quality of the plans is measured according to some “cost function”, such as an equation that combines total flight time and total fuel consumption.
In the course of their research, MIT scientists realized that the vast majority of the strategies of the 36 students participating in the study could be described using a formal language called linear temporal logic , which in turn could be be used to add restrictions to the specifications of the problem.
Using this language, they achieved an average improvement of the algorithms (in the resolution of numerical problems) of 13%; in the problems of flight planning and satellite positioning, of 16%; and in temporary problems, up to 12%.
The MIT researchers are now working on natural language processing techniques to make the system fully automatic, so that it converts users’ free descriptions of their own high-level strategies into a linear temporal logic, without human intervention.
At the end of January, they presented their advances in the International Conference on Automated Planning and Scheduling held at Carnegie Mellon University in Pittsburgh (USA).
More human inspiration
In the USA there is another recent project, from the American national intelligence agency IARPA and commissioned from Harvard University, to study the neural connections of the human brain, to apply them to the design of computer systems capable of interpreting, analyzing and learning like humans. This project aims to improve learning algorithms.
They are also trying to introduce into algorithms the ability to understand human humor, in this case, through Artificial Intelligence. In this line, a research team from Virginia Tech (USA), has created a trained algorithm to predict accurately when an image is funny and when it is not. The tool can also detect the most comical element of the scene and replace it to cause the opposite effect.