MAP-Elites can be problematic in face of uncertainty because:
- individuals can be unexpectedly lucky
- the behavior space can be hard to estimate and result in misplacing individuals.
Some mitigation techniques have been explored, e.g in (Justesen, Risi, and Mouret 2019) and this paper is about introducing another way of dealing with noisy domains without using sampling.
Here the main idea is to replace the MAP-elites grid by a “deep grid” with another dimension. This other dimension is used to store a population of individuals instead of a single individual for each elite. For each mutation, a single individual from the population is selected.
Experiments show that deep grids are a data efficient extension of MAP-elites which enable reducing uncertainty on the behavior descriptors.
- Flageat, Manon, and Antoine Cully. July 1, 2020. "Fast and Stable MAP-Elites in Noisy Domains Using Deep Grids". Artificial Life Conference Proceedings 32 (July). MIT Press:273–82. DOI.
- Justesen, Niels, Sebastian Risi, and Jean-Baptiste Mouret. July 13, 2019. “MAP-Elites for Noisy Domains by Adaptive Sampling”. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, 121–22. GECCO ’19. Prague, Czech Republic: Association for Computing Machinery. DOI.