Achieving Long-Term Progress in Competitive Co-Evolution

We illustrate how co-evolutionary experiments involving simulated predator and prey robots can lead to long-term global progress, i.e. can produce robots displaying progressively better performance against both competitors of current and previous generations. This is obtained by exposing evolving robots to well-differentiated competitors, by preserving individuals displaying good performance against hard to handle competitors, and by discarding opportunistic individuals that perform poorly against the other competitors of the current generation. The accumulation of variations producing general progress for more than 50,000 generations leads to the evolution of sophisticated behavioral capabilities and enable evolved robots to outperform robots evolved with simpler methods.

Publication type: 
Contributo in volume
Author or Creator: 
Simione, Luca
Nolfi, Stefano
Source: 
Proceedings of IEEE Symposium Series on Computational Intelligence, edited by D. Foegel and P. Bonissone (Eds.), pp. 855–862, 2017
info:cnr-pdr/source/autori:Simione, Luca; Nolfi, Stefano/titolo:Achieving Long-Term Progress in Competitive Co-Evolution/titolo_volume:Proceedings of IEEE Symposium Series on Computational Intelligence/curatori_volume:D. Foegel and P. Bonissone (Eds.)/edi
Date: 
2017
Resource Identifier: 
http://www.cnr.it/prodotto/i/396090
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8280898
Language: 
Eng
ISTC Author: 
Stefano Nolfi's picture
Real name: