Papers of the day   All papers

The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities

Comments

Jamie Bartlett: This is from an Atlantic article, about this paper: https://arxiv.org/pdf/1803.03453.pdf, which is VERY interesting - it's a collection of anecdotes of accidental creativity in algorithms. (Includes this good tic-tac-toe story, also mentioned in @TomChivers recent book). https://t.co/cokPaHYs71

3 replies, 322 likes


hardmaru: Found this recent paper by Tom Everitt and Marcus Hutter that looks at the topic of RL agents “cheating” from an AI Safety perspective. Worth a look! Paper https://arxiv.org/abs/1908.04734 Blog https://link.medium.com/yX1b3UXERZ https://t.co/cak1IUTAlM

3 replies, 121 likes


Janelle Shane: I notice people are retweeting this thread again - if you find this stuff as fascinating as I do, I definitely have a book for you: https://aiweirdness.com/books

1 replies, 29 likes


Jean-Baptiste Mouret: @AndrewM_Webb @hardmaru We observe this kind of things a lot when evolving robots or robot controllers / morphology. See https://arxiv.org/abs/1803.03453 for many examples.

0 replies, 13 likes


jonathanstray: I was lucky enough to be introduced to reward hacking when I saw Karl Sims' 1994 "Evolving Virtual Creatures" paper presented at SIGGRAPH. He showed examples of creatures cheating in hilarious ways, such as by falling over to "move" a long distance. https://arxiv.org/pdf/1803.03453.pdf https://t.co/QcgoZv4m84

1 replies, 11 likes


Physics Today: Joel Lehman et al point out that sometimes with ML/AI, the algorithms are too smart for their own good, such as deciding that crashing a plane into the ground is the most efficient way of landing it. H/T @JamieJBartlett https://arxiv.org/abs/1803.03453

1 replies, 6 likes


Machine Learning Failures: Neural networks evolved to classify edible and poisonous mushrooms took advantage of the data being presented in alternating order, and didn't actually learn features of the input images. (from https://arxiv.org/pdf/1803.03453.pdf)

0 replies, 5 likes


eigenrobot: @politicalmath @jaspar :) https://twitter.com/JanelleCShane/status/984809679040598016

1 replies, 4 likes


stoey: @GarethJennings3 @CherylRofer @luke_j_obrien Anyone wanting to give AI/Machine Learning a decision making position should be required to read this paper first: https://arxiv.org/pdf/1803.03453.pdf Nice summary thread of it here: https://twitter.com/JanelleCShane/status/984824347201232896

0 replies, 4 likes


Kevin J. Brennan: A perfect example of how the metrics you choose distort the outcomes you get.

0 replies, 2 likes


Machine Learning Failures: A virtual creature meant to evolve swimming strategies began traveling at unrealistic speeds after learning that it could generate more energy by twitching. (from https://arxiv.org/pdf/1803.03453.pdf)

0 replies, 1 likes


will: @hikari_no_yume Oh, I absolutely *love* these stories. You might be interested in this: https://arxiv.org/abs/1803.03453

0 replies, 1 likes


Henrik Korsgaard: AI must read IMO The Surprising Creativity of Digital Evolution: A Collection of Anecdotesfrom the Evolutionary Computation and Artificial Life Research Communities https://arxiv.org/pdf/1803.03453.pdf "creative" in research perhaps -- destructive and highly problematic in production

1 replies, 1 likes


Machine Learning Failures: A player discovers that it can “win” tic-tac-toe by making invalid moves, causing opponents to run out of memory, crash, and forfeit the match. (from https://arxiv.org/pdf/1803.03453.pdf)

0 replies, 1 likes


Liminal Warmth 🌸: Mostly what this thread tells me is that we need fewer bugs in the physics of our robot training simulators.

0 replies, 1 likes


Alasdair Allan: An intriguing paper, "A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities," https://arxiv.org/pdf/1803.03453.pdf. #MachineLearning

1 replies, 0 likes


101 144 141 155 040 105 154 153 165 163: @MoralHazardPay @antimule @LightningShade0 this is especially true because machine learning/evolutionary computation systems have a nasty habit of doing things like crashing the physics engine to give themselves more juice or inducing memory bombs to beat opponents https://arxiv.org/pdf/1803.03453.pdf

2 replies, 0 likes


Content

Found on Aug 19 2019 at https://arxiv.org/pdf/1803.03453.pdf

PDF content of a computer science paper: The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities