The primary annual conference on evolutionary computation, GECCO 2017, took place in Berlin in July. As always, it was an exciting event with much more to see and do than there was time. The conference has grown to over 600 participants, with 21 workshops, 33 tutorials, and 13 paper tracks. A total of 464 papers were submitted, with full paper acceptance rate of 39%. Here are some of our highlights:
Scaling black-box modeling with interaction graphs: Sometimes called gray-box modeling, this approach allows incorporating information about the domain that makes it possible to scale to a very large number (e.g. 1M) of variables, as shown by Chicano et al.
Open-ended evolution: Keynote by Drew Purves and Chrisantha Fernando pointed out the importance of the environment in facilitating evolution of complexity. While open-ended evolution is still challenging, their PathNet approach is taking a step in this direction by allocating different paths in a neural network to new tasks. A paper by Brant and Stanley suggested that minimal criterion coevolution of problems and solutions, e.g. mazes that can be solved by at least one agent, and agents that can solve at least one maze, could lead to open-ended complexity, because it does not require designing fitness or novelty functions that likely limit the complexity that can evolve. Hod Lipson’s keynote demonstrated another aspect of such coevolution: it can discover simple and elegant solutions, i.e. coevolution has a regularizing effect.
Deep neural network architecture optimization: Several papers proposed methods for evolving deep neural networks. The key currently is to limit the search to a space of known useful primitives. Khadka et al. showed that adding a memory block similar to that in Neural Turing Machine improves sequence processing memories significantly. Suganuma et al. showed how Cartesian Genetic Programming that’s focused on a few types can discover powerful convolutional network architectures even with limited computational resources.
Killer applications of evolution: As usual, the Human Competitive results competition (Humies) showcased a number of applications where evolution exceeds human design. The winning entry, by Harper et al, showed how causal models can be evolved to explain quantum correlations. Along the same lines, the best-paper award in real-world applications, by Yu et al., showed how EA can optimize the cuts in 3D printing to minimize necessary support structures. Spirited discussions emerged at the conference on applications of evolutionary algorithms that would further uniquely demonstrate its power, and would also potentially capture the imagination of the more general scientific community. Suggestions included automated programming and program repair, game agents in complex games such as Starcraft, general video game playing with neuroevolution, and open-ended evolution of behaviors in 2D and 3D physics simulations. Perhaps we will see some of these applications next year…looking forward to it!
The Sentient team participated by giving a tutorial on neuroevolution and with papers on Ascend (which received a Bronze award in the Humies) and on novelty-search stepping stones (which was nominated for a Best-Paper Award in Complex Systems). Sentient was also one of the sponsors of the conference; our booth showcased recent research by Sentient seen here: