EPS 2015 Agenda/Schedule

2015 Evolution of Physical Systems Workshop

Schedule and Agenda



Welcoming Remarks

John Rieffel, Union College



Roles of Model-Free Experiments in Evolutionary Robotics

Fumiya Iida and Luzius Brodbeck, Department of Engineering, University of Cambridge

The recent progress of robotics and IT technologies has enabled us to explore more flexible physically evolving systems. In our laboratory, for example, we have been taking advantage of affordable robotic manipulators and other automation technologies such as gluing machines and computer vision analysis, for the purpose of model-free evolution of physical robots [Brodbeck, et al. 2015].  Our platform is capable of handling cubic robotic modules, and gluing them in variations of morphological configurations such that constructed structures could exhibit diverse system-environment interactions such as dynamic crawling locomotion. The interactions can be also visually analyzed automatically, and the fitness evaluation can be used to improve morphological and controller designs in the real-world setup without any simulation models. Because most of the processes are automated, we are so far able to a large number of physical robots within a relatively short period of time (in the scale of hundreds of robots in a few weeks). Though the automation of robot construction and analyses of them is an interesting challenge by itself, we have been repeatedly questioned why simulated robot evolution is not sufficient. In this presentation, we would like to discuss a few distinctively important reasons why physical experiments of robot evolution are necessary, i.e. how the specifics of the real world drives evolution with interesting dynamics; how the real world implementation determines the level of abstraction; and how the details of real-world implementation can be compared to biological evolution in nature.


Evolving Computational Solutions in Matter

Julian Miller, University of York

Natural Evolution has been exploiting the physical properties of matter since life first appeared on earth.  Evolution-in-materio (EIM) attempts to follow this example by programming matter so that computational problems can be solved.  The beauty of this approach is that artificial evolution may be able to utilize
unknown physical effects to solve computational problems. This methodology is currently being undertaken in a European research project called NASCENCE: Nanoscale Engineering for Novel Computation using Evolution. We show how a variety of solutions to computational problems have been evolved using mixtures of carbon nanotubes and polymers at room temperature and also with gold nanoparticles at temperatures less than one Kelvin.



Going from {0, 1} to [0, 1]

Eivind Samuelsen and Kyrre Glette Department of Informatics University of Oslo

When working with physical systems, one has to deal with variance in performance measurements. The real world is filled with noise, imperfections and chaotic behavior that introduce unwanted measurement error. It is common to measure performance over long time intervals or average a large number of measurements in order to filter out this variance. This makes experiments take a long time, and may prohibit some kinds of experiments altogether. We argue that research into smarter, more efficient use of measurement data may significantly reduce the time needed for each complete test. This would of course improve evaluation budgets of existing evolutionary algorithms on physical systems, and may also help realize applications such as rapid on-line gait learning and adaption.

At present, immense computing power is available to us, even when carried on-board freely moving robots or similar systems. We suggest using this computing power to apply statistical methods like resampling methods, regression and hypothesis testing. By collecting more data in the same time frame, and using statistics to analyze it more thoroughly, we aim to reduce the time needed to do individual measurements, while maintaining sufficient measurement accuracy.

For example, while measurements often need to be done over time in- tervals of a certain minimum length, say one control system period, they might not need to be disjoint intervals in order to supply some useful sta- tistical information. By sampling often, and then computing the perfor- mance for many feasible subintervals, i.e. resampling the data, we can obtain a large sample of data that can inform us about the variance of our measurements, and can give us better performance estimates.

Having a data sample associated with each measurement, not just a scalar performance value, also enables us to use statistical methods to bet- ter compare noisy measurements. This gives us probability estimates of one measurement outperforming another instead of a binary true/false answer. Although these estimates may be inaccurate, they are likely to be no less misleading than a scalar comparison. Such estimates can also be used by algorithms to decide whether more evaluation time is needed before making a decision, or if an evaluation can be stopped early.

When comparisons result in probabilities instead of binary choices, we need to adopt our algorithms accordingly. Evolutionary algorithms al- ready employ randomness extensively, so the changes needed should not be too big or too uncomfortable. Binary tournaments can be won by at random based on the probability of being best. Definitive rankings can be replaced with most-likely rankings. Fitness proportional mechanisms can be replaced with or supplemented by probability-of-being-best proportional equivalents.

It is our opinion that gathering and making better use of more data, and making algorithms aware of measurement uncertainties, can reduce the time needed to make measurements sufficiently accurate and improve on what sufficiently accurate is. Experimental verification of these ideas is needed to quantify the effects a more advanced statistical approach to individual measurements.


15:30-16:00 Tea Break


Evolving gaits for physical robots with generative encoding: using single-unit pattern generators

Danesh Tarapore, Antoine Cully, and J.-B Mouret, UPMC



Evolution and Analysis of Morphological Computation in a Physical Tensegrity Robot

John Rieffel, Union College



Panel Discussion and Wrap-Up


Skip to toolbar