Evolving AI-Life by Natural Selection (Alastair Channon'sArtificial Life) A:link{text-decoration:none} A:visited{text-decoration:none} Background | Research | Publications | Software | Research-Related | Teaching | Contact The Evolutionary Emergenceof Increasingly Intelligent Behaviours via Computational NaturalSelectionA new approach to creating intelligence, rooted in Artificial Life andNatural Selection rather than traditional AI.  Research Context(Background) Simulated evolution is well established as a computationaloptimization methodology. However, this methodology has so farconcentrated on evolution towards pre-specified goals: a route whichcannot generate anything like the unbounded evolution or the diversityand complexity of structures that we observe in nature. In “On theOrigin of Species” Darwin emphasized the difference between thestruggle between organisms for limited resources (biotic competition)and the struggle against features such as drought, of the non-livingphysical environment (abiotic competition). Biotic competition, heargued, has been the cause of sustained evolutionary progress. Neo-Darwinism, which adds Mendelian heredity andpost-Mendeliangenetic theory, has clarified the nature and origins of species tothe extent that we can carry out evolutionary experiments withinartificial systems such as computer simulations and robotics.However, within the fields of artificial evolution (includingevolutionary/genetic algorithms/programming/computing and sub-fieldsof artificial life, adaptive behaviour and digital biota), work todate uses only abiotic competition, with very few exceptions. Most ofDarwin's theory would seem to have been ignored. Where bioticcompetition has been used, serious problems of evolvability can beidentified. The main focus is now on the generation of asystem that exhibits unbounded evolution. Bedau and Packard'sevolutionary activity statistics provide a basis for testing for thisand have already been applied to a number of both artificial andnatural selection systems, including my own. From the successes andfailures of these tests we are beginning to extend the list of knownrequirements for unbounded evolution. Beyond this, biologists will wantto use such systems to draw generic conclusions about evolution andengineers will want to evolve solutions that we can find uses for. My Research The emerging field of computational natural selection is anexciting area of artificial evolution that deals with the generationand analysis of open-ended evolutionary systems, primarily with theaims of overcoming the severe scaling problems exhibited by today'sevolutionary algorithms, and evolving computational intelligence[10-13]. In the computational natural selectionparadigm, thephenotype to fitness mapping is an emergent property of the evolvingenvironment and competition is biotic rather than abiotic. Geb is the only closedartificialsystem to have passed the statistical “Artificial Life Test” forunbounded evolutionary dynamics [7] and my refined(harder to pass)form of that test [1]. Earth's biosphere (throughits fossil records)is the only other system to have passed this test, in either form,although a number of Artificial Life systems have been evaluated.This is a very significant result: potentially a second example ofunbounded evolution, from which we can begin to draw generalizedconclusions about open-ended evolution that were not possible givenjust the real-world example. The creation of a system capable ofpassing the test had been identified by Bedau, Snyder and Packard as“among the very highest priorities of the field of artificiallife”. The significant refinements made to the test (including mymethod of component activity normalization) have been found by Stoutand Spector to be crucial in resisting attempts to achieve aclassification of unbounded dynamics in “intuitively unlifelike”systems. Extending this research is key to both our understanding ofevolution and our ability to replicate its full power to generateemergent processes and structures, including complex and intelligentones. The next logical step is to enable the observation of evolvedbehaviours as they emerge. This can best be achieved through theevolution of both agents’ morphologies and controllers within asimulated physical environment, which has the added advantage ofproviding an open range of low-level actions. As a first step, preliminary work on such a system, but withsimpler, fitness-function based evolution, has been carried out: wehave presented a ‘modified replication’ [3-5] ofKarl Sims’work on the evolution of articulated artificial creatures inphysically realistic 3D environments. Our system was the first todemonstrate comparable results to those of Sims, and did so usingstandard McCulloch & Pitts’ neurons rather than ad hocelements, so as not to provide problem-specific a priori knowledge. Publications E. Robinson, T. Ellis and A. D. Channon,``Neuroevolution of Agents Capable of Reactive and Deliberative Behaviours in Novel and Dynamic Environments,''accepted to appear in Advances in Artificial Life: Proceedings of the Ninth European Conference on Artificial Life (ECAL 2007), Springer Lecture Notes in Computer Science (LNCS) volume ????, in the Lecture Notes in Artificial Intelligence (LNAI) subseries.This paper presents evolved artificial neural controllers that solve tasksrequiring deliberative behaviours: tasks that cannot be solved by reactivemechanisms alone and which would traditionally have their solutionsformulated in terms of search-based planning. Two very different neuralnetworks are used: one that controls high-level deliberative behaviours,such as the selection of sub-goals, and one that provides reactive andnavigational capabilities. Animats controlled by a hybrid of thesenetwork architectures are evolved in novel and dynamic environments,on increasingly complex versions of an example problem. The resultsdemonstrate, for the first time ever, incremental neuro-evolutionarylearning on such tasks. A. D.Channon,``Unboundedevolutionary dynamics in a system of agents that activelyprocess and transform their environment,'' Genetic Programming andEvolvable Machines, vol. 7, no. 3, pp. 253–281, 2006.This paper presents significant improvements to the establishedstatistical "ALife Test" for unbounded evolutionary dynamics.Its contribution is in making the revised test (which the systemfrom publication 8 still passes) well-grounded even for long-termunbounded evolution in artificial systems, through the first ever methodof computing individual genes' adaptive ('normalized') evolutionaryactivities. In their paper "Validation of evolutionary activity metricsfor long-term evolutionary dynamics", Stout and Spector attemptedto "break" the test by achieving unbounded dynamics in "intuitivelyunlifelike" systems. They concluded that Channon's method of activitynormalization is of particular importance to the test's robustnessagainst such attempts. T.Miconi andA. D. Channon,``TheN-Strikes-Out Algorithm: A Steady-State Algorithm forCoevolution,'' in Proceedings of the 2006 IEEE Congress onEvolutionary Computation (CEC 2006), IEEE World Congress onComputational Intelligence (WCCI 2006), Vancouver, (G.G. Yen et al.,eds.), pp. 1639–1646, IEEE Press, 2006. T.Miconi andA. D.Channon,``Analysingcoevolution among artificial 3D creatures,'' inProceedings of the 7th International Conference on ArtificialEvolution (Evolution Artificielle 2005): Revised Selected Papers,Lille, (E.G. Talbi et al., eds.), pp. 167–178, Springer, 2006. Avolume of the LNCS Series.This paper presents new accomplishments in the coevolution of neurallycontrolled agents, and introduces improved methods of coevolutionaryanalysis. The experiments reported, on the coevolution of physicallysimulated articulated creatures, are the first to demonstrate realisticco-adapted behaviours using general purpose neurons. The previous needfor ad hoc (problem-specific) neurons was a barrier to the long-termevolution of new, emergent behaviours. Novel behaviours are identifiedusing an improved coevolutionary analysis method that is both moreinformative and an order of magnitude cheaper than the original.Finally, individuals are cross-validated between evolutionary runs,in an improved procedure for evaluating global performance. T.Miconi andA. D.Channon,``AnImproved System for Artificial Creatures Evolution,'' inProceedings of the Tenth International Conference on the Simulationand Synthesis of Living Systems (ALife X), Bloomington, Indiana (L.M.Rocha et al., eds.), pp. 255–261, MIT Press, 2006. T.Miconi andA. D.Channon,``Avirtual creatures model for studies in artificial evolution,'' inProceedings of the 2005 IEEE Congress on Evolutionary Computation(CEC 2005), Edinburgh (D. Corne et al., eds.), volume 1, pp. 565–572,IEEE Press, 2005. A. D.Channon,``Improvingand still passing the ALife test: Component-normalisedactivity statistics classify evolution in Geb as unbounded,'' inProceedings of Artificial Life VIII, Sydney (R. K. Standish, M.A. Bedau and H. A. Abbass, eds.), (Cambridge, MA), pp. 173–181,MIT Press, 2003. Instructionsfor replicating the runs discussed in this paper. A. D.Channon,``Passingthe ALife test: Activity statistics classify evolution inGeb as unbounded,'' in Advances in Artificial Life: Proceedings ofthe Sixth European Conference on Artificial Life (ECAL2001), Prague(J. Kelemen and P. Sosik, eds.), (Heidelberg), pp. 417–426,Springer Verlag, 2001. A volume of the LNCS/LNAI Series. This paper presents the first ever closed artificial system to pass theestablished statistical "ALife Test" for unbounded evolutionary dynamics:an achievement identified by Bedau, Snyder and Packard as "among the veryhighest priorities of the field of artificial life". Earth's biosphere(through fossil-record databases) is the only other system to have passed,although many have been evaluated. This is a very significant result: wecan now begin to draw generalized conclusions about open-ended evolutionthat were previously impossible given just the real-world example.It significantly advances our ability to generate emergent processesand structures, including complex and intelligent ones. A. D.Channon,``EvolutionaryEmergence: The Struggle for Existence in ArtificialBiota,'' PhD thesis, Department of Electronics and ComputerScience,University of Southampton, 2001. A. D.Channon, ``Threeevolvability requirements for open ended evolution,'' inArtificial Life VII Workshop Proceedings (C. C. Maley and E.Boudreau, eds.), (Portland, OR), pp. 39–40, 2000. A. D.Channon andR. I. Damper, ``Towardsthe evolutionary emergence ofincreasingly complex advantageous behaviours,'' InternationalJournalof Systems Science, special issue on Emergent Properties of ComplexSystems, vol. 31, no. 7, pp. 843–860, 2000. A. D.Channon andR. I. Damper, ``The evolutionary emergence of sociallyintelligent agents,'' in Socially Situated Intelligence: a workshopheld at SAB'98, University of Zurich Technical Report (B. Edmondsand K. Dautenhahn, eds.), (Zurich), pp. 41–49, 1998. A. D.Channon andR. I. Damper, ``Perpetuatingevolutionary emergence,'' in FromAnimals to Animats 5: Proceedings of the Fifth InternationalConference on Simulation of Adaptive Behavior (SAB98), Zurich(R. Pfeifer, B. Blumberg, J. A. Meyer, andS. Wilson,eds.), (Cambridge, MA), pp. 534–539, MIT Press, 1998. A. D.Channon andR. I. Damper, ``Evolvingnovel behaviors via naturalselection,'' in Proceedings of Artificial Life VI, Los Angeles(C. Adami, R. Belew, H. Kitano, and C. Taylor,eds.), (Cambridge, MA), pp. 384–388, MIT Press, 1998. A. D.Channon andR. I. Damper, ``The artificial evolution of real intelligence bynatural selection.'' Published on the web site of and posterpresented at the Fourth European Conference on Artificial Life(ECAL97), Brighton, 1997. A. D.Channon,``TheEvolutionary Emergence route to Artificial Intelligence.'' MScthesis, School of Cognitive and Computing Sciences, University ofSussex, 1996. Software GebVersion7c5 for Unix (Linux/SunOS/IRIX/HP-UX/...) & MS-Windows Geb is an artificial world containing organismswhichevolve by natural selection. The papers above provide the bestdescription of Geb. I recommend reading at least Perpetuatingevolutionary emergence before trying to make sense of the sourcecode, and that Passingthe ALife test is the next paper you read. |
|
