Evolving Learnable Neural Networks Under Changing Environments with Various Rates of Inheritance of Acquired Characters: Comparison of Darwinian and Lamarckian Evolution

The processes of adaptation in natural organisms consist of two complementary phases: learning, occurring within each individual's lifetime, and evolution, occurring over successive generations of the population. In this article, we study the relationship between learning and evolution in a simple abstract model, where neural networks capable of learning are evolved using genetic algorithms (GAs). Individuals try to maximize their life energy by learning certain rules that distinguish between two groups of materials: food and poison. The connective weights of individuals' neural networks undergo modification, that is, certain characters will be acquired, through their lifetime learning. By setting various rates for the heritability of acquired characters, which is a motive force of Lamarckian evolution, we observe adaptational processes of populations over successive generations. Paying particular attention to behaviors under changing environments, we show the following results. Populations with lower rates of heritability not only show more stable behavior against environmental changes, but also maintain greater adaptability with respect to such changing environments. Consequently, the population with zero heritability, that is, the Darwinian population, attains the highest level of adaptation to dynamic environments.