Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments

Ability to continuously learn and adapt from limited experience in environments is an important milestone on the path towards intelligence. In this paper, we cast the problem of continuous into the learning-to-learn framework. We develop a simple-based meta-learning algorithm suitable for adaptation in dynamically and adversarial scenarios. Additionally, we design a new multi-agent environment, RoboSumo, and define iterated adaptation games for various aspects of continuous adaptation strategies. We demonstrate meta-learning enables significantly more efficient adaptation than baselines in the few-shot regime. Our experiments with a population of that learn and compete suggest that meta-learners are the fittest.