Towards Explainable AI Algorithms via Fitness Landscape Analysis in Evolutionary Computation

There lies a motivation in investigating how well meta-heuristic search methods like Genetic Algorithms perform when used alongside illumination algorithms like MAP-Elites to confidently select the best candidate solutions from the solution space. From these 'best' candidate solutions, the focus is then on how these solutions can be explained using XAI algorithms, and how well XAI algorithms can be represented using Fitness Landscapes. There is research to support the use of EC and MAP-Elites together [1] and suggests that their use can be explored to help better accommodate user preferences. There is also research that supports the use of XAI and landscape analysis, with one study finding that the landscape was neutral overall. A neutral landscape can indicate redundant features exist within model configurations, meaning that the end analysis is not as refined as it could be [3]. Therefore, it is important to ensure that the models only contain the most pertinent features that are needed for the transformation especially when there is a reliance on the outcome of the model.
Fitness landscapes are a useful way to keep track of the model fitness generated by XAI algorithms, and XAI methods could also be used in fitness landscapes to reinforce the strength of the EC algorithm and the trust from the end-users. The interplay between these terms opens up a lot of potential to research new areas where this could be applied, for instance this could be implemented in healthcare diagnosis tools, the decision could be supported by a number of factors pertinent to that illness, which can be achieved using an XAI method known as Counterfactual Analysis [2]. The fitness of this model can be shown on a fitness landscape, where this model can be generated across 10 simulated runs. In each run, the number of best candidate solutions (models) could be selected using MAP-Elites. These solutions will have an associated fitness score and can be plotted on a fitness landscape to reinforce how strong the model is across each of the independent simulations, using something like decision trees to determine if the landscape analysis is good or bad. This is just one possible scenario, other scenarios can be modelled and experimented with to discover the potential of integrating these technologies. Currently, there exists a number of XAI tools that try to bridge this gap in understanding for end-users[4]. The interest will be in seeing how well these things work together for a set of different problem domains, where the requirements for each end-user will differ.