Request to readers. + Results from maximisation of accelerometer reading

Fascinating to see people from all over the world reading this. Would be very interested to hear from you, to know who you are. Do leave a comment :)

I pray to the reinvented sacredness that NAO will evolve behaviour routines that look impressive that allow it to maximise the aggregate score over its accelerometer over the course of the trial. In the meantime....


Evolving fitness functions (Co-evolution of tests and solutions) 

Co-evolutionary variance guides physical experimentation in

evolutionary system identification (2005)

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.134.7830

The above paper removes tests in which the models have too much disagreement, i.e. where fitness on this test is very low on all the models. That is, very difficult games can be removed because the fitness scored on these games is very low. 

also in the same conceptual ballpark:


http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=4983086

http://www.demo.cs.brandeis.edu/papers/ficici_gecco03.pdf

http://eprints.soton.ac.uk/265255/1/gecco.pdf

http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=489152

http://creativemachines.cornell.edu/papers/TEC08_Schmidt.pdf

http://www.dagstuhl.de/Materials/Files/08/08351/08351.SWM4.Slides.pdf

http://download.bioon.com.cn/view/upload/month_0807/20080731_9f738b4390be637eb03aE2coCuyuHBL2.attach.pdf

http://155.97.32.9/~plutynsk/FundamentalTheorem

[7] Bongard, J. C. and H. Lipson. Automated robot function recovery after unanticipated failure or environmental change using a minimum of hardware trials. In Proceedings of The 2004 NASA/DoD Conference on Evolvable Hardware, pp. 169–176, Seattle, WA, 2004.
[8] Lipson, Hod and Josh Bongard, An Exploration- Estimation Algorithm for Synthesis and Analysis of Engi- neering Systems using Minimal Physical Testing, Proceed- ings of the ASME Design Automation Conference (DAC04), Salt Lake City, UT, Sept. 2004.

How to encode a rich space of motor actions with a small number of parameters? 

Not the nicest rise in fitness. I think there is a bloat issue with molecule sizes and the deletion operator needs to be per locus probability and not singular. 

See video for some of the final solutions. 

The accelerometer changes during a posturing are shown below

The middle component shows the values at rest, and the left and right parts during a posturing over which fitness is assessed. Note, the values of the blue and green readings are higher (the red reading slightly lower). Overall, the value is higher during the posturing, and thus, that posture is selected for. The posture is achieved by pushing off with one foot to try to move the chest accelerometer to a higher value. No more complex motor plan to achieve higher accelerometer values was discovered however, and this is attributable I think to the rather boring way in which atoms encode motor postures perhaps?

Methods for more complex motor encoding of joint motion

1. Looks like we can use Kuramoto oscillators (Phil's stuff). Each oscillator is a joint and the joints are coupled. There are regimes where the coupling is synchronous, and asynchronous. See http://www.sussex.ac.uk/Users/rm262/papers/CEC10.pdf

2. We can implement a Liquid State Machine and allow the readouts to be joint angles. Readout properties can be evolved to produce a rich variety of dynamical processes in the joint outputs. 

Lets start with a LSM atom, i.e. a LSM in a single atom, with parmaterized injections of inputs to the Liquid and paramaterized perceptron readout outputs. 

3.CPGs for hierarchical reinforcement learning. http://staff.science.uva.nl/~whiteson/pubs/snelicdl11.pdf

Long run results 

Results from a long run in which the z + x + y accelerometer reading over the assessment is maximised. Possible a weird thing to try to do!

Videos show behaviour. The negative numbers above are where the robot is on its front, and the positive when it is on its back.

Turning onto its front:

Turning onto its back:

Vera has some suggestions to refine and understand this result. The first is to visualise the motor molecules during the course of the run, the other is to produce a more interesting fitness function/game than just accumulating accelerometer scores over the whole trial.