Weeks 4 + 5

Events:

• Evolution Gym won’t work!!!
• NEW members of lab
• Problem with the fitness function not recognizing the material
• Bigger project: evolutionary robotics
• Professor Soros sick.

Because of the shift in the lab, evolution gym not working, and professor Soros being sick, it was a very research-y week. I ended up researching more evolutionary robotics rather than EAs.

Reading List

• On the Entanglement between Evolvability and Fitness: an Experimental Study on Voxel-based Soft Robots, by Dr. Soros
• On a Feasible–Inf easible–Infeasible T easible Two-Population (FI-2P opulation (FI-2Pop) Genetic op) Genetic Algorithm for Constrained Optimization: Distance Tracing and no Free Lunch, by Steven O. Kimbrough
• Surrogate Infeasible Fitness Acquirement FI-2Pop for Procedural Content Generation, by Gallotta et al.

Group research

Since more people joined the lab, Dr. Soros organized events where we all would meet up and discuss our overarching topic she is focusing on: creative means for studying feasibility in evolutionary life. Since her research focuses on game design and game research, this week, out reading was:

• IEE Conference on Games 2021 Keynote

PROBLEM THIS WEEK: Evolution not working on operating systems

This week, we tried to download the Evolution Gym software, but unfortunately the packages would not load and the python files would not work. Given this, Dr. Soros worked with me to debug my issue, trying things such as:

• Use a VM with different operating systems
• Stack Overflow
• Using different computers all together
• Contacting the people of Evolution Gym
• Recruiting more people to figure out the issue Nothing, unfortunately worked. Thus, we may have to change platforms. This, however, will not change the trajectory of our research, as their other other voxel-based platforms for soft robots.

Thus, much of my work was focused on research and integrating myself with the larger research group.

Define more terms

Neural networks

Memory! Complex digital systems that take the form of human-modeled neural or brain connections. They can represent different “neurons” or “types” or neurons.

The Reality Gap

Solutions generated in simulation may not always work in “reality” on tangible robots; hence, there is a “gap” in what is generateable vs. concienvable.

Reduce the reality gap: Transferability functions, encourage development of robust controllers (“envelope of noise”, Jakobi 1995), encourage reactivity of controllers (?), add online adaptation abilities to improve controller robustness.

Selective pressure

External agents that influence the survivability for an agent in an environment. External agents may have more impact that the encodings.

Complexification

A theoretical basis for whether robot complexity is defined through “robustness” or neural networks.

Reinforcement learning

Similar to ER, they both deal with agents obtaining rewards by interacting in their envriinments to achieve the optimal outcome, with the reward/outcome being the fitness value. On the other hand, ERs only use the global value for the outcome and not for state-action pairs.

Open-ended evolution

The aspect of evolution that is constantly capable of morphological and behavioral changes and innovations. Qualitites: increasing diversity and complexity, continuous adaptation.

Transferability functions

Function that prediicts the limits of simulations themselves. Transfers dozens of controllers during the evolutionary process to track both performance and simulation.

Genotype-phenotype map

A map for neural networks that links genes to their traits. The mappings seek to reproduce considency, regularity, modularity, and hierarchy from natural systems.

Evolvability

The ability for a population to “evolve” over time, quantified on a group capacity, not on the individual.

The Baldwin Effect

(in relation to learning and evolution conjointly): suggests that learning may facilitate evolution to create a selective pressure, which in turn makes it easier to discover optimal behaviors. Explores the interaction between learning and evolving and whether they lead to the same outcome.

Environment-Driven Evolutionary Robots

he reliance of robots to develop based on non-enginerained “DNA” but reaction to the world around. A limitation is that the environment changes often, which is difficult for robots to anticipate and adapt to uncertainty.