Self-Adaptive Mutation Operators for Genetic Neural Networks in Survival Analysis

Artificial neural networks (ANN) are used within the medical eld of survival analysis to rank patients according to their risk group. To evaluate how well the ranking was conducted, it is common to obtain the concordance index error (c-index). It has been shown that ANNs can be trained directly on the c-index with the use of genetic algorithms (GA). The GA evolution of an ANN is controlled by a set of operators, which in turn are governed by hyperparameters. These hyperparameters are usually static and set to generally accepted good values, or optimised through a grid search for each specific data set. In this article, adaptive and self-adaptive techniques are introduced to the hyperparameters that governs the mutation operators. It is shown that while training on the c-index, it is possible to simultaneously revise the mutation width. Furthermore, it is found that these techniques can be generalised to other aspects of mutation, and show promising results of the possibility of a self-adaptive genetic algorithm, independent of initialisation and data set.How to train your AI: repiton, repetion, repetition Artificial intelligence(AI) has been on the rise within the medical field worldwide in the last five years. From image recognition software to nd skin cancer, to big data analysis creating treatment suggestions, these machines are becoming more common and promising day by day. In the short term, medical professionals will be able to request an AI analysis as easily as they would a blood test. In the long term, we could see machine diagnostics dominating entire medical fields. As AI becomes accepted throughout society, the shortcomings become more apparent. The most central problem is the lack of general artificial intelligence. There are single-problem oriented programs that do one thing well but fails when applied to other problems. No need to fear an AI take-over any time soon then. Instead, we are definitely at a point in time were human and computer intelligence can join forces and achieve great things that was previously not possible. This study used machine learning combined with selfadaptive evolutionary training. To understand these concepts, imagine you try and score a three pointer in basketball. You shoot, you miss, you try again. Every time you shoot, you correct your shot, shooting harder or softer, and thus improve the longer you keep going. A machine learns similar tasks in the same way, by repeating and learning from its mistakes. But learning is more complicated than that. You don't just aim higher or lower, you might jump to get a better angle, or change how you move your hands. The same goes for the machine. Each time it tries to solve a problem, it evolves and gets better, changing depending on what change has the biggest positive impact. So how do you know how much to change and when? The machine must be allowed to learn and change in every way it needs for it to solve the problem at hand. By applying these principles, this work presents an attempt to develop a program that could solve a set of similar but varied medical problems, resulting in SAGA, or Self-Adaptive Genetic Algorithm. The program is difierent from previous similar programs in that it changes depending on the problem, and it does so without a human intervention. SAGA managed to find solutions that were better than previous static programs, relying less on the user of the program, as well as the data it was given. The learning techniques that SAGA explored should be of interest to the medical field, as it makes the use of machine learning more accessible, but also to anyone interested in the pursuit of a truly general and independent AI