In the NeurIPS 2018 Artificial Intelligence for Prosthetics challenge, participants were tasked with building a controller for a musculoskeletal model with a goal of matching a given time-varying velocity vector. Top participants were invited to describe their algorithms. In this work, we describe the challenge and present thirteen solutions that used deep reinforcement learning approaches. Many solutions use similar relaxations and heuristics, such as reward shaping, frame skipping, discretization of the action space, symmetry, and policy blending. However, each team implemented different modifications of the known algorithms by, for example, dividing the task into subtasks, learning low-level control, or by incorporating expert knowledge and using imitation learning.

Abstract. Evolution in Materio (EIM) exploits properties of physical systems to compute. "Designs" are evolved instead of a traditional top down design approach. Computation is a product of the state(s) of the material and input data. Evolution manipulates physical processes by stimulating materials assessed in situ. A hardware software platform designed for EIM experimentation is presented. The platform with features especially for EIM is described together with demonstration experiments using carbon nanotubes in a thick film placed on micro electrode arrays.

No abstract

Abstract. Evolution-in-materio (EIM) is a method that uses artificial evolution to exploit the properties of physical matter to solve computational problems without requiring a detailed understanding of such properties. EIM has so far been applied to very few computational problems. We show that using a purpose-built hardware platform called Mecobo, it is possible to evolve voltages and signals applied to physical materials to solve machine learning classification problems. This is the first time that EIM has been applied to such problems. We evaluate the approach on two standard datasets: Lenses and Iris. Comparing our technique with a well-known software-based evolutionary method indicates that EIM performs reasonably well. We suggest that EIM offers a promising new direction for evolutionary computation.

Abstract-Evolution-in-materio (EIM) is a method that uses artificial evolution to exploit properties of materials to solve computational problems without requiring a detailed understanding of such properties. In this paper, we show that using a purpose-built hardware platform called Mecobo, it is possible to evolve voltages and signals applied to physical materials to solve computational problems. We demonstrate for the first time that this methodology can be applied to function optimization. We evaluate the approach on 23 function optimization benchmarks and in some cases results come very close to the global optimum or even surpass those provided by a well-known software-based evolutionary approach. This indicates that EIM has promise and further investigations would be fruitful.