Artificial Life Lecture 19

1. Artificial Life Lecture 19 Summing up! Bits on PSO, Developmental Robotics. Spiking NNs Survey of Artificial Life topics covered And Alife skills Recent EASy research Personal prejudiced list of hot research topics Time for questions? Artificial Life lecture 19

2. A bit on PSO -- Particle Swarm Optimisation • Rather like a GA Population searches in N-dimensional Genotype space – a PSO Swarm searches in N-dimensional (real-valued) Sequence Space, with some flocking algorithm: • Each particle in swarm has a velocity and a fitness • Each velocity changes with an acceleration that depends (noisily) on weighted contributions from both Personal-best-so-far and Global-best-so-far • So the accelerations are rather like mutations, cross-influence from Global-best-so-far a bit like recombination Artificial Life lecture 19

3. PSO Animation Copied with permission from Maurice Clerc, PSO Mini Tutorial See via www.particleswarm.info Different flavours to PSO, implicitly have something like variable mutation rates (cf Evolution Strategies), competitive with GAs on some problems Artificial Life lecture 19

4. PSO update equations (typical, can vary) For each particlei (1): Vi = wVi(t-1) + c1r1(pBesti-pLoci) + c2r2(gBesti-pLoci) (2): pLoci = pLoci(t-1) + Vi w: weight/momentum V: N-dim Velocity vector i: Current particle index t: Current iteration c1, c2: Weighting constants r1, r2: Stochastic variables [0 < r < 1], gBest: N-dim Global ‘population’ best-so-far pBest: N-dim Particle ‘personal’ best-so-far pLoc: N-dim Particle current location Artificial Life lecture 19

5. Developmental Robotics Possibly now better known as “Epigenetic Robotics”. From an Epigenetic Robotics Conference call:- Epigenetic systems … a prolonged developmental process … cognitive and perceptual structures emerge … interaction of an embodied system with a physical and social environment. Goals include: (1) understanding biological systems by the interdisciplinary integration of social and engineering sciences and (2) enabling robots and other artificial systems to autonomously develop skills for new environments (instead of programming them to solve problems in fixed environments). Psychological theory and empirical evidence is being used to inform epigenetic robotic models, and these models can be used as theoretical tools to make experimental predictions in developmental psychology. Artificial Life lecture 19

6. Dynamics of Development Cf. Luc Berthouze -- Dynamics of Development spring option Behavioural plasticity in animals and humans, multi-timescale processes in agent/environment interaction, self-organized value-dependent development, dynamical and autonomous systems modelling for developmental systems, evolution of developmental strategies, aspects of neural development, aspects of social development, aging and dynamics of developmental disorders Artificial Life lecture 19

7. Spiking Neural Networks Recordings from real NNs often show spike trains. Is this no more than a biological way of transmitting Hi or Lo signals, according to the spike frequency? – or are the details of the timing doing more than that? Artificial Life lecture 19

8. Spiking Different motives: e.g. modelling how an individual neuron spikes Hodgkin-Huxley model, etc. Or hypothesising radically different methods of neural processing – e,g. synfire chains, cf Abeles Artificial Life lecture 19

9. Or spiking Artificial NNs e.g. Floreano, D., Epars, Y., Zufferey, J.-C. and Mattiussi, C. (2006)  Evolution of Spiking Neural Circuits in Autonomous Mobile Robots. International Journal of Intelligent Systems, 21(9) p1005-1024. Neuron model is a simple integrate-and-fire model with leakage and refractory period. Possible advantages? May be easy to implement cheaply and crudely in hardware? Artificial Life lecture 19

10. Alife Topics covered - Evolutionary • Evolution, Evolutionary Algorithms • Co-evolution • Evolution of communication • Development, L-Systems, G->P mappings • Fitness Landscapes • Neutral Networks • Information and Life and Evolution • Tierra and Avida • GP and Classifier Systems Artificial Life lecture 19

11. Alife topics – DS and robotics, and beyond • Dynamical Systems approach to cognition • Braitenberg vehicles • CTRNNs • Evolutionary Robotics • Passive Dynamic Walking • CAs and RBNs • Models of Genetic Regulatory Systems • Morphogenesis, L-Systems • Homeostasis • Gaia Theory, Daisyworld • Artificial Chemistry • Autocatalysis, autopoiesis Artificial Life lecture 19

12. Alife Skills • Programming a GA • Microbial GA • Programming a CTRNN • Programming a robot • (coming next term) ODE physics engine • How to prepare a project proposal • (…hopefully…) how to carry through a project • Research. In Spring term I shall give a seminar on “How to apply for a PhD/DPhil” • Main points will be: you have to find a research topic and a supervisor • UK/Europe, apply in Spring, N. American system different • … jobs outside academia .. Artificial Life lecture 19

13. From the CCNR web-page Artificial Life lecture 19

14. Artificial Life lecture 19

15. Artificial Life lecture 19

16. Artificial Life lecture 19

17. Artificial Life lecture 19

18. Artificial Life lecture 19

19. Artificial Life lecture 19

20. Artificial Life lecture 19

21. Artificial Life lecture 19

22. Artificial Life lecture 19

23. Artificial Life lecture 19

24. Allister Furey – using Evolutionary Robotics to control kites for Wind Energy http://www.afurey.com/ Artificial Life lecture 19

25. Bill Bigge’s Programmable Springs Compliant actuators – regulating force rather than position http://www.informatics.sussex.ac.uk/users/wb23/PGS/Videos.html Artificial Life lecture 19

26. Hot topics:- Passive Dynamic Walking This is a very embodied, dynamical systems approach. So far PDWs have just gone down gentle slopes under gravity. There is scope for adding small amounts of power input for more general walking. Matt Williamson’s work, associated with Brooks’ COG project, using coupled oscillators, seems a very promising lead that could be applied here. Eric Vaughan took this a lot further in his thesis – still room for much more work. Artificial Life lecture 19

27. Neutral Networks Wide open for research. Barnett’s work shows that in a formally defined class of (binary) fitness landscapes full of NNs in a particular fashion, best strategy is a population of size 1+1, with a fixed number of mutations based on getting expected proportion that are neutral as close as possible to 1/e = 37%. Adrian Thompson’s hardware evolution supports this Extensions: noisy fitness evaluations ? Real valued genotypes? Artificial Life lecture 19

28. Gaia/Maximum entropy? Relationship between Daisyworld models of homeostasis, and Thermodynamic ideas that :- “Systems ‘try to organise themselves’ to produce entropy as fast as possible”. Cf. Kay and Schneider, ‘4th Law of Thermodynamics’. “Organised systems can dissipate junk entropy faster, and (… speculation…) Life does it better than anything else and hence should naturally occur (… given the right circumstances)” MEP Maximum Entropy Production Principle Artificial Life lecture 19

29. Homeostasis Homeostasis seems central to definitions/the core of life/cognition (cf Autopoiesis) As also does Jim Stone’s work on (… roughly speaking…) what perceptual systems (eg ANNs) have to do in order to sift out higher-level invariants from all the noise. Roughly:- (1) output of a system should not fluctuate wildy (just echoing the noise is pointless) But (2) Should not stay still – it should fluctuate over the long term if it is to reflect real things happening in the world. Artificial Life lecture 19

30. Economy So, Jim Stone points out that outputs of of perceptual systems should minimise (Short Term Variance) divided by (Long Term Variance) Minimise STV/LTV where short/long refer to appropriate time scales. This is one way of ensuring that an output neuron is “earning its keep”. Artificial Life lecture 19

31. Rein Control • Rein Control (originally Manfred Clynes) appears to be a really simple unifying principle to explain • Vanilla Daisyworld • McDonald-Gibson’s DW variant • Homeostasis generally. • Immense scope for research … … Artificial Life lecture 19

32. Adaptive Text Entry methods The need for easier methods for text entry on mobile phones, on iPods, PDAs etc. Particularly with iPhone, iTouch. EASy style approach, adaptive interface. Fitts’ law. One finger moving can generate ~14 bits info/sec, hence in theory ~10 characters of English text/sec, ~120 wpm. Artificial Life lecture 19

33. Autonomous Glider Adding minimal sensors and a relatively simple Braitenberg-like control architecture to a model glider. Aim: to get autonomous flight gaining lift/power through ridge-soaring. Minimal optical-horizon sensor, 3 DoF accelerometer, use of optic flow. Current EU joint proposal Artificial Life lecture 19

34. Maintain position relative to a ridge – how? Compare the optical horizon with the gravity horizon -- should keep you somewhere on the red line. But you also need to keep within appropriate distances from ridge, on that red line. Use optic flow? Unstructured texture of ground – interesting issues. Artificial Life lecture 19

35. Summer research projects Please remember the range of possibilities covered here (..and suggest more…) when it comes to planning your main summer projects. Please also remember that the robots in the ASL are in general available for your own independent projects. There will be a talk in Spring Term on ‘How to apply for a PhD/DPhil place’ Artificial Life lecture 19

36. The End … time for more discussion … ?? REMEMBER: Feedback on courses Via Study Direct !!! Artificial Life lecture 19