ARC-Solver10

By: Somayyeh Gholami & Mehran Kazeminia

Abstraction and Reasoning Challenge :

• The ARC challenge was first launched on "Kaggle" in February 2020. Mr. François Chollet and his colleagues had prepared and arranged the corpus of tasks as well as the rules and regulations of the competition. We and the rest of the participants tried hard to get a good score, but it did not happen and most of the participants did not succeed. Various types of neural networks, different machine learning algorithms and other efforts such as using the Z3 programming language did not have good results.

• After the challenge time was over, most of the participants published their notebooks. The successes were relatively limited, but different views can be seen in these notebooks. Of course, we started our work again and using our new ideas and the ideas in the public notebooks, we managed to solve at least thirty percent of the tasks.

• In 2023, the ARC challenge was again held at "Lab42" and we managed to get the best score in that competition. Our method was a combination of "Sklearn tree" and "Genetic Algorithm" as well as several solvers that solve tasks with specific characteristics. Finally, we combined our results with the results of the icecuber notebook. Our notebook was called "Solver 7".

Explanation for "Solver 10" method :

• ARC is more than a competition for us and our main goal is not just to get a better score. Rather, it is interesting for us to understand how and with what process the machine takes the initial steps of "Abstraction and Reasoning" and we are interested in being a part of this story. That's why we designed and implemented "Solver 10" and are still developing it.

Advantages of the "Solver 10" method :

Last Explanation :

• Machines typically go through three steps to answer ARC tasks (as well as humans to answer intelligence questions). In the first stage, all the information of a task, including inputs and outputs, is checked, and based on these observations, several "hypotheses" are considered to perform the task correctly. In the second stage, these hypotheses are tested, so that the correct theory is chosen. In the third stage, the best theory is implemented for ARC, or for human intelligence questions, the correct option is selected among several options.

• Humans are much better than machines in the first stage. But in the second and third stages, the matter can be exactly the opposite. That is, if all the basic keys are given to the machine, the power and speed of the machine will be thousands of times human. And this great power in the second and third stages can overcome the weaknesses of the first stage.