Artificial life based on neural network animals - "Merks" - evolving in a rich environment
Keywords: Artificial life, Soft Articial Life, Artificial Neural Networks, Deep neural networks, Genetic Algorithms, Evolution, Natural Selection, Open-Ended Evolution, Swarm Intelligence, Emergence
The world of Merks has been largely inspired from Geb [1], but is presented in a more generic, modular way, which I hope will help a broad re-use of the code and spreading of the experimentation.
Merks are little artifical animals living in an artificial environment, and each of them have a Neural Network and a Genome. A Merk has a sense of the surrounding environment and takes actions thanks to the Neural Network. The Neural Network itself is build according to its binary Genome at the time of birth. By design, each Merk has only innate response and can not learn from experience in its lifetime. However, from generation to generation the Merk populations undergo evolutionary changes. With natural selection, the Merks develop an increasingly rich and adapted behaviour. The Genome of Merks evolve thanks to pairwise "sexual" reproduction - the genome of a new Merk is created from merging two parent Merks, and also with random mutations.
The world can be diverse and present challenges to the Merks, that foster interesting evolutionary changes and behaviour. My belief is that it is possible to see sophisticated population dynamics, intra-species social behaviour, inter-species symbiosis, predation or cohabitation, and offer a sandbox for what should be the dream and goal of Artificial Life Research: open-ended evolution.
My second belief is that this model is strong enough to produce interesting result by putting Merks in new, challenging environments. My invitation to you is to take on this project yourself, and build new environment rules that produce surprising results. Change the physics, change the reproduction system if you see fit. Experiment and come back to share your discoveries ! Let us enter together the area of Life 2.0.
[1] A. D. Channon, Improving and still passing the ALife test: Component-normalised activity statistics classify evolution in Geb as unbounded, in Proceedings of Artificial Life VIII, Sydney (R. K. Standish, M. A. Bedau and H. A. Abbass, eds.), (Cambridge, MA), pp. 173–181, MIT Press, 2003. Instructions for replicating the runs discussed in this paper.