Evolution in Cognition Workshop @GECCO 2018

Kyoto, Japan

July 15 12:50 to 14:30

Program

Introduction
- 12:50 - 12:55, Introduction to the Workshop
Invited talks
- 12:55-13:20 Emergence of Social Behaviors in Reinforcement Learning Agents, Stefan Elfwing
- 13:20-13:45 Deep Neuroevolution, Joel Lehman
- 13:45-14:10 Building Creative Robots by Means of Evolutionary Algorithms, Stéphane Doncieux
Short talks (accepted submissions):
- 14:10-14:20 Meta Learning by the Baldwin Effect, Chrisantha Fernando
- 14:20-14:30 The Flouted Naming Game: Contentions and Conventions in Culture, Harold de Vladar

Workshop Description

Evolution by natural selection has shaped life over billions of years leading to the emergence of complex organism capable of exceptional cognitive abilities. These natural evolutionary processes have inspired the development of Evolutionary Algorithms (EAs), which are optimization algorithms widely popular due to their efficiency and robustness. Beyond their ability to optimize, EAs have also proven to be creative and efficient at generating innovative solutions to novel problems. The combination of these two abilities makes them a tool of choice for the resolution of complex problems.

There is evidence that the principle of selection on variation is at play in the human brain, as proposed in Changeux’s and Edelman’s models of Neuronal Darwinism, and more recently properly reformulated in the theory of the Neuronal Replicators. Consequently, the idea of an interaction between evolutionary processes and cognition over physiological time scales has been gaining some traction. Since the development of human cognition requires years of maturation, it can be expected that artificial cognitive agents will also require months if not years of learning and adaptation. It is in this context that the optimizing and creative abilities of EAs could become an ideal framework that complement, aid in understanding, and facilitate the implementation of cognitive processes. Additionally, a better understanding of how evolution can be implemented as part of an artificial cognitive architecture can lead to new insights into cognition in humans and other animals.

The goals of the workshop are to depict the current state of the art of evolution in cognition and to sketch the main challenges and future directions. In particular, we aim at bringing together the different theoretical and empirical approaches that can potentially contribute to the understanding of how evolution and cognition can act together in an algorithmic way in order to solve complex problems. In this workshop we welcome approaches that contribute to an improved understanding of evolution in cognition using robotic agents, in silico computation as well as mathematical models.

Keywords: Evolutionary Computation, Evolution, Cognition, Darwinian Neurodynamics, Neuronal Darwisnism, Robotics.

Submissions and Deadlines

Accepted submissions:

Original: experimental work, position papers as well as overviews of author's recent work are all welcome. These submissions will be included in the proceedings and be presented during the workshop.
Abstracts submitted as posters to GECCO 2018: the authors of selected abstracts will be given a talk. The abstract won’t be included in the proceedings.
Hot-of-the-press: authors can submit a recently published work. Selected articles will be given a talk.

Workshop papers must be submitted using the GECCO submission system this year. After login, the authors need to select the Workshop Paper submission form. In the form, the authors must select the workshop they are submitting to. To see a sample of the Workshop Paper submission form go to GECCO's submission site and chose Sample Submission Forms.

Submitted papers must not exceed 8 pages (shorter submissions are also encouraged) and are required to be in compliance with the GECCO 2018 Papers Submission Instructions. It is recommended to use the same templates as the papers submitted to the main tracks. It is not required to remove the author information.

All accepted papers will be presented at the corresponding workshop and appear in the GECCO Conference Companion Proceedings.

Deadlines

~~March, 27th~~ April 3, 2018: Papers Submission Deadline
April, 10th, 2018: Notification of Acceptance
April, 24th, 2018: Camera-Ready Paper Submission
July 15 or 16th, 2018: Workshop

Invited Speakers

We are pleased to announce that the following internationally recognized researchers will present their work related to the workshop topic and propose their own views on the related questions. Each invited speaker will give a talk and also participate in a panel discussion at the end of the workshop.

Stefan Elfwing

ATR Computational Neuroscience Labs

Japan

Emergence of social behaviors in reinforcement learning agents

How adaptive agents with self interests can acquire social behaviors including cooperation and communication is a fascinating open issue. We report our relevant findings, including learning how, what and whether to communicate and evolution of alternative mating strategies.

Joel Lehman

Uber AI

United States

Deep Neuroevolution

A recurring lesson in the deep learning era is the value of revisiting old neural network algorithms in the light of massive data, computation, and network size: The old can be made new, often revealing substantial gains in performance and in the difficulty of problems that can be tackled. We are beginning to see that this pattern not only holds for supervised backpropagation and reinforcement learning algorithms like Q-learning, but extends also to neuroevolution algorithms. Work out of Open AI with a form of evolutionary strategies first demonstrated that evolution can compete with more traditional deep reinforcement learning algorithms on their own terms. This talk reviews recent work out of Uber AI Labs inspired by and building upon that result. Our results demonstrate that, surprisingly, vanilla genetic algorithms too can compete with deep reinforcement learning, even establishing a state-of-the-art result in one Atari game and a highly compressible network encoding. Complementarily, we find that gradient information can enable safer mutations that make deep neuroevolution more efficient. Finally, other of our work examines the neuroevolution algorithm applied by OpenAI, aiming to understand its behavior more deeply, and also to push it further, by extending it with novelty search. The conclusion is that deep neuroevolution is an exciting and promising frontier of research, one still waiting to fully benefit from the many advances and insights accumulated by the evolutionary computation community over the years.

Stéphane Doncieux

Sorbonne University

France

Building creative robots by means of evolutionary algorithms

Programming a robot to deal with an environment as diverse as our everyday environment remains a challenge, whereas humans can easily deal with it. One of the reasons may be their creativity, that allow them to discover solutions that go beyond their initial knowledge. Todays robots are preprogrammed or, when they learn, they are limited by the primitives or demonstrations they have been provided with. To build more creative robots and in the frame of the DREAM project, we have proposed a developmental approach in which the evolutionary principles of variation and selection, that neuroscientists think to be at play in human creativity, help creating repertoires of skills that may allow a robot to face situations unknown to its programmer. We will present the current results of the project.