In recent years reinforcement learning (RL) has received a lot of attention thanks to its performance and ability to address complex tasks. At the same time, multiple recent papers, notably work from OpenAI, have shown that evolution strategies (ES) can be competitive with standard RL algorithms on some problems while being simpler and more scalable. Similar results were obtained by researchers from Uber, this time using a gradient-free genetic algorithm (GA) to train deep neural networks on complex control tasks. Moreover, recent research in the field of evolutionary algorithms (EA) has led to the development of algorithms like Novelty Search and Quality Diversity, capable of efficiently addressing complex exploration problems and finding a wealth of different policies while improving the external reward (QD) or without relying on any reward at all (NS). All these results and developments have sparked a strong renewed interest in such population-based computational approaches.
Nevertheless, even if EAs can perform well on hard exploration problems they still suffer from low sample efficiency. This limitation is less present in RL methods, notably because of sample reuse, while on the contrary they struggle with hard exploration settings. The complementary characteristics of RL algorithms and EAs have pushed researchers to explore new approaches merging the two in order to harness their respective strengths while avoiding their shortcomings.
Some recent papers already demonstrate that the interaction between these two fields can lead to very promising results. We believe that this is a nascent field where new methods can be developed to address problems like sparse and deceptive rewards, open-ended learning, and sample efficiency, while expanding the range of applicability of such approaches.
With the Evolutionary Reinforcement Learning Workshop, we want to highlight this new field currently developing while proposing an outlet for the two communities (RL and EA) to present new applications and ideas and discuss past and new challenges.
Authors are encouraged to submit original research articles, case studies, reviews, position papers, and theoretical papers within the following topics of interest:
Evolutionary reinforcement learning
Evolution strategies
Population-based methods for policy search
Neuroevolution
Hard exploration and sparse reward problems
Deceptive reward
Novelty and diversity search methods
Divergent search
Sample-efficient direct policy search
Intrinsic motivation, curiosity
Building or designing behaviour characterizations
Meta-learning, hierarchical learning
Evolutionary AutoML
Open-ended learning