Evolutionary Robotics and its application to time travel
Evolutionary Robotics and its application to time travel
Our recent paper "Simple Robotic Simulation of Perception, and Time Travel" should eventually appear on our website and it can be found in our blog of March 21st,2011.
Precise graphical plots of the movement of these machines is still being made, to create a family of curves under various circumstances and for various configurations, as well as for varying datings of internal robot time.
One following approach could entail the deciding of which of these configurations can give the most optimum results. Simple mathematical neural network theory has already failed to give any definite real promise in other connections. So many obvious examples have shown its basic failings, the typical case, perhaps, being in the case of calculations for derivative instruments in the stock market. That is an interesting case as, whilst simple mathematical neural network theory should conceivably be able to take into account most contingent circumstances, just as weather predictions can be expected to lead to some positive results from very simple macro measurements including wind speed, temperature, barometric pressure and so on, perhaps over a very wide (or even global) area without a lot of micro understanding, leaving aside the inevitable problems of chaos theory and the like, some results of a plausible nature may be obtained for weather predictions. But generally we can still think of Lorenz (even allowing for the fact with Cray computers etc. we have gone some way past Lorenz's Royal McBee computer), or on a practical basis be aware that most people have little faith in newspaper weather predictions, for example. But for the stock market, even agreeing that macro parameters should give perhaps at least a rough indication, generally speaking it is no easier than predicting winners of a horse race or results within a cricket game, simply because of fudging by the jockeys or players in ways which the macro parameters might well be, in essence, incapable of prediction. Even in these simple cases, it can be hard or impossible to predict which banker or jockey or cricket player will do the fudging, or indeed whether it may be some entirely extraneous factor quite outside the scope of the state of play and even outside the metric of the statistical measurements.
This does not prevent us from using evolutionary robotics, together with neural networks (Nolfi, 1994).
The simple message at the moment is that we are not currently trying to 'predict the future' results, but rather to establish a working McTaggart A series format for description. Our earlier working experiments used the real results and experimental methods of Gregory (1970), where he used visual techniques to observe astronomic observations - which related to signals from objects now in our own past (Yates, 1984). We evolved the techniques of Gregory so they would work with electroencephalographic measurements in our own laboratories and with the results of Kornhuber et al. etc. and our own readings, and achieved a working time machine.
Therefore at a later point we may use, at the Institute for Fundamental Studies, the Information Field Theory of Torsten Ensslin (Ensslin, 2009) and the help of Feynman diagrams to carry though some of the mathematical integration difficulties.
It is hoped to be able to duplicate the effect of A series time travel, not necessarily with a live human being as we have already done in the way of a psychological experiment, but in a crude way in a very simple system by the use of very basic swarm robotics. Changizi has already pointed out that a simple and somewhat similar B series effect is possible even with existing biomeasurements.
Miglino et al (2008) describe a simple new approach in Evolutionary Robotics according to which human breeders can become involved in the evolutionary process. They use their simple "Breedbots" which is not very different to our own NXT robots (Yates, 2011) or indeed to the usual Braitenberg type 2 robots. That is almost like the creation of an experimental philosophy or psychology interface with non human participants and as such, also has some theoretical interest.
So now we need to redesign a testing plan to see how accurately our robots can obtain type A McTaggart behaviour. This is not the same as building robots which will to all intents and purposes predict the future, although these robots should have the qualities of relevance and experience to Type A behaviour.
The process is likely to be to set up a course which may contain puzzles or obstacles and to see which variations in the robots need to be altered, using evolutionary neural nets, to create robots with the response that we require. We may judge performance by our own subjective judgement as to which robot qualities were successful, or allow a neural network or other mathematical techniques to decide. Tests may need a swarm of robots or the compilation of results by a sequence of modified robots.
Winfield (2011) admits very fairly and reasonably that "Right now we just don't know how to design a system that produces complex overall behaviours from a group of simple agents" so as usual the Institute for Fundamental Studies will start as simply as possible, hoping for results from just one modified model, perhaps with derived videos obtained using the model in several measured configurations - producing a single video with perhaps six approximate clones appearing within it, following different paths which have been measured using an isolated robot to obtain responses as varied by model modifications. The appropriate sets of rules can then be added to each of the six slightly different clones. Then the relative successes of the clones can be assessed from viewing the video which will show how a simulation using all the six in the same video succeeds. To be quite fair, I have looked at many of the swarm robot videos on Youtube, today April 15th, 2011 and they do seem to be unsatisfactory and unfortunately sometimes little more than child's play, with no real advantages. Clearly it is a difficult task to show or prove anything significant at all, but here at the Institute for Fundamental Studies we at least have purpose, method, and current knowledge and we are doing much better then many others in our opinion.
The above work should form the basis of a further note.
References
Esslin T.A.,Frommert M., Kitaura F.S.,(2009), "Information field theory for cosmological perturbation reconstruction and nonlinear signal analysis",Phys. Rev. D 80, 105005
Gregory R.L. (1970),"The Intelligent Eye", Appendix B, 170 et seq., Weidenfeld & Nicholson, London
Miglino O.,Gigliotta O.,Ponticorvo M, Lund H.H.,(2008), "Human Breeders for Evolving Robots", Artificial Life and Robotics , vol. 13, no. 1, pp. 1-4
Nolfi S.,Floreano D.,Miglino O., Mondada F., (1994),"How to evolve autonomous robots: different approaches in evolutionary robotics", in R.A. Brooks, P. Maes eds., Proceedings of the IV International Workshop on Artificial Life, Cambridge, MA, MIT Press
Winfield A.,(2011), http://www.ias.uwe.ac.uk/~a-winfie/ ,
http://www.newscientist.com/article/dn13244-shapeshifting-robot-forms-from-magnetic-swarm.html
Yates J.,(1984) Patent Number:GB2051465 Publication date:1981-01-14. I also mention and apply Gott's comment to this patent in http://philpapers.org/archive/YATASO.1.pdf
Our recent paper "Simple Robotic Simulation of Perception, and Time Travel" should eventually appear on our website and it can be found in our blog of March 21st,2011.
Precise graphical plots of the movement of these machines is still being made, to create a family of curves under various circumstances and for various configurations, as well as for varying datings of internal robot time.
One following approach could entail the deciding of which of these configurations can give the most optimum results. Simple mathematical neural network theory has already failed to give any definite real promise in other connections. So many obvious examples have shown its basic failings, the typical case, perhaps, being in the case of calculations for derivative instruments in the stock market. That is an interesting case as, whilst simple mathematical neural network theory should conceivably be able to take into account most contingent circumstances, just as weather predictions can be expected to lead to some positive results from very simple macro measurements including wind speed, temperature, barometric pressure and so on, perhaps over a very wide (or even global) area without a lot of micro understanding, leaving aside the inevitable problems of chaos theory and the like, some results of a plausible nature may be obtained for weather predictions. But generally we can still think of Lorenz (even allowing for the fact with Cray computers etc. we have gone some way past Lorenz's Royal McBee computer), or on a practical basis be aware that most people have little faith in newspaper weather predictions, for example. But for the stock market, even agreeing that macro parameters should give perhaps at least a rough indication, generally speaking it is no easier than predicting winners of a horse race or results within a cricket game, simply because of fudging by the jockeys or players in ways which the macro parameters might well be, in essence, incapable of prediction. Even in these simple cases, it can be hard or impossible to predict which banker or jockey or cricket player will do the fudging, or indeed whether it may be some entirely extraneous factor quite outside the scope of the state of play and even outside the metric of the statistical measurements.
This does not prevent us from using evolutionary robotics, together with neural networks (Nolfi, 1994).
The simple message at the moment is that we are not currently trying to 'predict the future' results, but rather to establish a working McTaggart A series format for description. Our earlier working experiments used the real results and experimental methods of Gregory (1970), where he used visual techniques to observe astronomic observations - which related to signals from objects now in our own past (Yates, 1984). We evolved the techniques of Gregory so they would work with electroencephalographic measurements in our own laboratories and with the results of Kornhuber et al. etc. and our own readings, and achieved a working time machine.
Therefore at a later point we may use, at the Institute for Fundamental Studies, the Information Field Theory of Torsten Ensslin (Ensslin, 2009) and the help of Feynman diagrams to carry though some of the mathematical integration difficulties.
It is hoped to be able to duplicate the effect of A series time travel, not necessarily with a live human being as we have already done in the way of a psychological experiment, but in a crude way in a very simple system by the use of very basic swarm robotics. Changizi has already pointed out that a simple and somewhat similar B series effect is possible even with existing biomeasurements.
Miglino et al (2008) describe a simple new approach in Evolutionary Robotics according to which human breeders can become involved in the evolutionary process. They use their simple "Breedbots" which is not very different to our own NXT robots (Yates, 2011) or indeed to the usual Braitenberg type 2 robots. That is almost like the creation of an experimental philosophy or psychology interface with non human participants and as such, also has some theoretical interest.
So now we need to redesign a testing plan to see how accurately our robots can obtain type A McTaggart behaviour. This is not the same as building robots which will to all intents and purposes predict the future, although these robots should have the qualities of relevance and experience to Type A behaviour.
The process is likely to be to set up a course which may contain puzzles or obstacles and to see which variations in the robots need to be altered, using evolutionary neural nets, to create robots with the response that we require. We may judge performance by our own subjective judgement as to which robot qualities were successful, or allow a neural network or other mathematical techniques to decide. Tests may need a swarm of robots or the compilation of results by a sequence of modified robots.
Winfield (2011) admits very fairly and reasonably that "Right now we just don't know how to design a system that produces complex overall behaviours from a group of simple agents" so as usual the Institute for Fundamental Studies will start as simply as possible, hoping for results from just one modified model, perhaps with derived videos obtained using the model in several measured configurations - producing a single video with perhaps six approximate clones appearing within it, following different paths which have been measured using an isolated robot to obtain responses as varied by model modifications. The appropriate sets of rules can then be added to each of the six slightly different clones. Then the relative successes of the clones can be assessed from viewing the video which will show how a simulation using all the six in the same video succeeds. To be quite fair, I have looked at many of the swarm robot videos on Youtube, today April 15th, 2011 and they do seem to be unsatisfactory and unfortunately sometimes little more than child's play, with no real advantages. Clearly it is a difficult task to show or prove anything significant at all, but here at the Institute for Fundamental Studies we at least have purpose, method, and current knowledge and we are doing much better then many others in our opinion.
The above work should form the basis of a further note.
References
Esslin T.A.,Frommert M., Kitaura F.S.,(2009), "Information field theory for cosmological perturbation reconstruction and nonlinear signal analysis",Phys. Rev. D 80, 105005
Gregory R.L. (1970),"The Intelligent Eye", Appendix B, 170 et seq., Weidenfeld & Nicholson, London
Miglino O.,Gigliotta O.,Ponticorvo M, Lund H.H.,(2008), "Human Breeders for Evolving Robots", Artificial Life and Robotics , vol. 13, no. 1, pp. 1-4
Nolfi S.,Floreano D.,Miglino O., Mondada F., (1994),"How to evolve autonomous robots: different approaches in evolutionary robotics", in R.A. Brooks, P. Maes eds., Proceedings of the IV International Workshop on Artificial Life, Cambridge, MA, MIT Press
Winfield A.,(2011), http://www.ias.uwe.ac.uk/~a-winfie/ ,
http://www.newscientist.com/article/dn13244-shapeshifting-robot-forms-from-magnetic-swarm.html
Yates J.,(1984) Patent Number:GB2051465 Publication date:1981-01-14. I also mention and apply Gott's comment to this patent in http://philpapers.org/archive/YATASO.1.pdf
posted by uv at 6:08 AM
0 Comments:
Post a Comment
Links to this post:
Create a Link
<< Home