arXiv2023
In recent years, studies in epigenetic inheritance in biological systems as well as studies on evolution in non-biological systems e.g., machine learning and robotics, have reopened the discussion of non-Darwinian methods of evolutionary optimization. In this paper, the three most prominent classical evolutionary strategies Lamarckian, Darwinian, and Baldwinian are implemented and compared in an agent-based simulation. The dynamics of optimization and learning are studied in constant as well as in dynamic environments, both as single evolutionary strategy populations and as mixed evolutionary strategy populations. The three different agent types are implemented as simple objects that need to gather resources to replicate. The agent fitness is defined through a bitstring match between the agent and the environment and to replicate each agent needs to gather sufficient resources. To make a fair comparison between the three different evolutionary methods, we can assume that both the ability to learn within an individual's lifetime and the learning process itself require additional resources compared to an evolutionary process without lifetime learning. When all three strategies coexist and we vary the replication costs for the different strategies we find the following general pattern: The evolutionary advantage decreases for Lamarckian and Baldwinian strategies as their learning costs increase and/or as their relative replication costs increase compared to the Darwinian strategies, and in most situations the Baldwinian strategy performs in between the Darwinian and Lamarckian strategies. Finally, as our investigations emphasize the superiority of the Lamarckian strategy in most cases with low learning cost, it is interesting to note that a corresponding biological inheritance mechanism did not evolve, while the Lamarckian strategy is currently extensively used in technology applications.