The Science of Evolutionary Intelligence
Besides clear application potential, there are clear scientific challenges to be met and boundaries to be pushed. The science of Evolutionary Intelligence lies in the understanding of how to build scalable and competent combinations of EC and ML algorithms and to understand the associated limits of these resulting algorithms, i.e., which problem classes can be effectively solved. This specifically concerns the question of how key structural features of problems such as decomposability, multi-modality, and hierarchy can be captured in efficiently representable and exploitable models. This has been a key driving force of the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) research line, with the envelope still being pushed today.
Another key challenge, on the side of XAI, is how to learn hierarchically decomposable explainable AI models efficiently and effectively, e.g., through functional encapsulation methods so that we can move beyond simple sets of equations being discovered and build complex, yet at every level human-understandable models.
On the neuro-evolution side, it is an open question how we can efficiently simultaneously search for the best macro structure of the architecture of deep neural networks as well as novel operations at a microlevel, to find innovative architectures well-suited for specific tasks beyond what can be found by only combining blocks of pre-defined operations.
While these are important challenges in their own right, an overarching grand challenge is how to bring the gears of EC and ML even closer together to realize (evolutionary) optimization algorithms that do not have to restart from scratch when solving a new problem but can learn from previous results (on similar problems) to be more efficient each time. This grand challenge can even be extended to the ML side, with the ultimate goal of achieving efficient, self-configuring, and self-improving AI technology that is well beyond the current state of AutoML that makes it much easier for practitioners to create state-of-the-art AI solutions within their environments.
