Assignment #16

Provide a short discussion of each of the assigned papers (listed under Course Materials). Below are some questions to think about.

Karl Sims

Questions

• This is the second paper Karl Sims' wrote on evolving virutal creatures. The first paper (which you can optionally read) does not involve explicit competition between agents. Instead, a bunch of candidate creatures are each simulated in their own separate environments, and fitness is assessed via an absolute measure, like distance moved. The creatures that move the farthest are selected for reproduction, resulting in creatures that can run or swim quickly. In Karl Sims' second paper (which you read), the simulation is a bit different. It pits one creature against another in the same physical environment and they play a symmetric two-player game. The creature that wins is selected for reproduction. What do you think is an advantage of this setting over the former? What's a possible disadvantage?
• Sims evaluates different design choices based on subjective judgment (how "interesting" is the resulting behavior). Can you think of more objective metrics? How could we decide if one design choice (e.g., tournament pairings) is better than another (e.g., random pairings)?

Hide and Seek

Questions

• The authors state that randomizing the environment during training was critical. Why do you think this is?
• The authors claim "there are no direct incentives for agents to interact with objects or to explore, but rather the emergent strategies are solely a result of the autocurriculum induced by multi-agent competition." Do you agree with this claim? What aspects of the environment, neural architecture, optimization algorithm, or experiments might "bake in" some of the behaviors we are seeing.
• Is the transfer learning evaluation compelling? If not, how should we measure emergent intelligence?

Upload a single PDF file through Stellar by Apr 16 at 10 am.