# ALife 2020

tags
ALife Conference, Artificial life

## Day 2

### Sara Walker, keynote Lecture

About what life means and how it can be defined from the point of view of physics/information theory, etc.

### Melanie Mitchell, keynote Lecture

This talk was very similar to another one I watched from Santa Fe Institute which promotes her book: Artificial Intelligence: A Guide for Thinking Humans.

It gives a detailed breakdown of why machine learning still fails at many tasks and might not be able to overcome them before we change the current way of doing it. The main takeaway for me is that Mitchell believes we will need to somehow build abstraction or conceptualization capabilities within our ML models to make them able to use “common sense” and understand things we consider as basic.

## Day 3

### Contributed session 1

#### Djordje Grbic, Sebastian Risi - Safe Reinforcement Learning through Meta-learned Instincts

See notes <&grbicSafeReinforcementLearning2020>.

#### Toby Howison, Josie Hughes, Fumiya Iida - Morphologically programming the interactions of V-shaped falling papers

Interesting talk about falling V-shaped papers. The range of exhibited behavior is surprising and is an instance of complex systems with a few controllable parameters but surprising dynamics.

#### Manon Flageat, Antoine Cully - Fast and stable MAP-Elites in noisy domains using deep grids

See notes <&flageatFastStableMAPElites2020>.

#### Julian F. Miller: Evolving developmental neural networks to solve multiple problems

This talk is about introducing a new model of neural network neuron that could have enough modularity to deal with multiple problems at once.

#### Peter Andras - Composition of Games as a Model for the Evolution of Social Institutions

Socials institutions can be seen as large games with rules, norms, and values. They are complex evolving systems starting form simple cooperation and evolving to complex things. The author applies a game theory framework to these systems.

#### James M. Borg, Matt Grove, Fiona Polack - Coloured noise time series as appropriate models for environmental variation in artificial evolutionary systems

Many biological time-series exhibit a type of colored noise. The authors use different noise functions in evolutionary models and observe that it profoundly affect the evolutionary response. This is interesting because it shows that the most often used white noise (Gaussian noise) might not be the best choice for evolution simulations.

### Contributed session 2

#### Randall D. Beer - An Integrated Perspective on the Constitutive and Interactive Dimensions of Autonomy

See notes <&beerIntegratedPerspectiveConstitutive2020>.

#### Mark Bedau - Dynamic natural kinds: Open-ended evolution in the joints of nature

The authors use word2vec on patents to identify technological feature spaces. The similarity between patent descriptions is used to measure their proximity. This offers and interesting perspective on the space of different technologies and their hierarchical and dynamical properties.

#### Acacia Ackles, Austin Ferguson, Connor Grady, Charles Ofria - Rank epistasis: A new model for analyzing epistatic interactions in the absence of quantifiable fitness interactions

This paper studies an epistasis model. The new proposed model is based on ranking loci of a genetic model based on their fitness. Then, after mutating a locus and computing the new ranking. By comparing statistical differences between the two rankings one can estimate the effect of the previously introduced mutation.

The authors used NK model to study the new metric in a relatively well studied environment. Sites with higher epistatic activity score higher according to their metric and this happens without any a priori information about how the sites interact with each other.

#### Thomas Helmuth, Edward Pantridge, Grace Woolson Lee Spector - Genetic Environment Sensitivity and Transfer Learning in Genetic Programming

This paper is about using genetic programming for program synthesis.

#### Mahi Luthra, Eduardo J. Izquierdo, Peter M. Todd - Cognition Evolves with the Emergence of Environmental Patchiness

Using a simple agent based simulation of animals and plants in a grid world, the authors observe spontaneous “patchiness” of the partition of plants. Cognition, defined as the ability perceptual strength of an agent and the stochasticity of its decisions increases and stabilizes in those simulations.

When cognition is not restricted, patches are larger and more visible, whereas for restricted cognition resource distribution is more uniform. There is a two way influence of both parameters on the evolution of patches and cognition.

#### Lauren Benso, Lauren Benson, Madhavun Candadai and Eduardo Izquierdo - Neural reuse in multifunctional circuits for control tasks

This talk is about studying a simple two-layers feedforward neural network evolved to solve three RL tasks at the same time. Then by introducing neuron lesions and their effect on the fitness, one is able to identify the useful neurons and their importance.

Many neurons are re-used across multiple tasks. With neural variance and mutual information within each neurons the authors are able to analyze the reuse of neurons.

The takeaways are :

• It seems relatively simple to obtain neuron reuse across tasks with simple evolutionary methods
• Successful neural networks on theses toy tasks appear to reuse their neurons

### Contributed session 3

#### Guillaume St-Onge, Antoine Allard, Laurent Hébert-Dufresne - Localization, bistability and optimal seeding of contagions on higher-order networks

In this talk, the author studies higher-order networks, which differently from simple networks, have interactions that can exist within groups of elements (instead of pairs). This allows studying contagion and spreading dynamics from a new angle.

#### Oskar Elek, Joseph N. Burchett, J. Xavier Prochaska and Angus G. Forbes - Monte Carlo Physarum Machine: Agent-based Model for Reconstructing Complex 3D Transport Networks

Cosmic web is diffuse, not only composed of galaxies but hot and cold gas and dark matter. It is very hard to capture from experimental data. Authors use the slime mold as a model for reconstructing this network. This mold is interesting because by using gradients of chemicals emitted by food and itself, it can construct near-optimal transport networks. Their model is also inspired from Sage Jensons’s cyber critters adapted to 3D. They use Monte-carlo sampling of directions around the gradient for moving each tiny agent which results in more natural-looking simulations.

This model reproduces the cosmic web in a relatively faithful way visually, and is robust thanks to its stochastic component.

### Poster Session

#### Matthew Dale - Designing Computational Substrates using Open-Ended Evolution

This poster is about characterizing the computational capabilities of reservoir computing systems. The author’s goal is to design computational substrates that have interesting properties when used as reservoir computing basis.

For this, the authors have to define a behavior space that they slice up into voxels. They use novelty search to explore that space and measure the quality of the possible behavior.

#### Elissa Cohen - Resilient Life: An Exploration of Perturbed Autopoietic Patterns in Conway’s Game of Life

See notes <&cikaResilientLifeExploration2020>.

## Day 6

### Contributed session 11

See notes.

See notes.

#### Merihan Alhafnawi, Sabine Hauert, Paul O’Dowd - Robotic Canvas: Interactive Painting onto Robot Swarms

Novelty search is used in simple maze and navigation tasks and different evolved and non evolved architecture are compared.

### Rebecca Kramer-Bottiglio keynote Lecture

This keynote was about using robotic skins to make soft and “natural”-looking robots. This allows to transform inert object into actuators or moving robots.

Kramer’s team is developing molecules and composite materials with multiple functionalities.