http:^^www.demo.cs.brandeis.edu
来自「This data set contains WWW-pages collect」· EDU 代码 · 共 148 行
EDU
148 行
Date: Wed, 20 Nov 1996 19:01:11 GMT
Server: Apache/1.1.1
Content-type: text/html
Content-length: 5223
Last-modified: Thu, 05 Sep 1996 15:47:43 GMT
<HTML><HEAD><TITLE>DEMO</TITLE></HEAD><BODY background="background.gif" text="#000020"LINK="#000050" VLINK="#0000f0" ALINK="#000050"><Center><!WA0><IMG SRC="http://www.demo.cs.brandeis.edu/manbutsmall.gif" hspace=10><h2> Dynamical & Evolutionary Machine Organization<br>Brandeis University</h2></center><p><!WA1><IMG SRC="http://www.cs.brandeis.edu/graphics/lines/color/line_colors_fade.gif"> <br><p><h2> DEMO attacks problems in agent cognition using complex machine organizations which are createdfrom simple components with minimal human design effort.We study recurrent neural networks, evolutionary computation,and dynamical systems as substrates. We build working systems to test our theories. DEMO is locatedin the back of Room 118Volen Center, next to the <!WA2><A href="http://www.cs.brandeis.edu/~agents">Interaction Lab</a>, but our entry is through the workstationcluster (Room 119) or the shop (117).</h2><p><!WA3><IMG SRC="http://www.cs.brandeis.edu/graphics/lines/color/line_colors_fade.gif"> <br><p><h2> Members of the Organization: <!WA4><A href="http://www.demo.cs.brandeis.edu/demofolk.gif">(& Photo)</a> <p><ul><!WA5><A href="http://www.cs.brandeis.edu/~pollack"> Jordan Pollack</a></h2><p><h3><!WA6><A href= "http://www.cs.brandeis.edu/~blair/home.html"> Alan Blair</a><br><p><!WA7><A href="http://www.cs.brandeis.edu/~hugues/hugues.html">Hugues Juille</a><br><p><!WA8><A href="http://www.cs.brandeis.edu/~sevan">Sevan Ficici</a><br><p><!WA9><A href="http://www.cs.brandeis.edu/~pablo/indexe.html"> Pablo Funes</a> <br><p><!WA10><A href="http://www.cs.brandeis.edu/~spitz/spitz.html"> Ayal Spitz</a> <br><p><!WA11><A href="http://www.cs.brandeis.edu/~hornby/index.html"> Gregory Hornby</a> <br><p><!WA12><A href="http://www.cs.brandeis.edu/~thecat/academic/"> Kerry Kurian</a> </h3></ul><p><!WA13><IMG SRC="http://www.cs.brandeis.edu/graphics/lines/color/line_colors_fade.gif"> <br><p><h2>Works in progress</h2><h3> Co-evolution of a Backgammon Player.</h3>One of the persistent themes in artificial life research is theuse of co-evolutionary arms races in the development of specific andcomplex behaviors. However, other than Sims' work on artificial robots,most of the work has attacked very simple games of prisoner's dilemma orpredator and prey. Following Tesauro's work on TD-Gammon, we used simplehill-climbing in a 4000 parameter feed-forward network to develop acompetitive backgammon evolution function. No back-propagation,reinforcement or temporal difference learning methods were employed.Instead we start with an initial champion of all zero weights and proceedsimply by playing the current champion network against a slightly mutatedchallenger and changing weights when the challenger wins. Evaluationproceeds by applying the network to all legal moves and choosing the movewith the highest evaluation. Our results show co-evolution to be apowerful machine learning method, even when coupled with simplehill-climbing, and that the surprising success of Tesauro's program hadmore to do with the co-evolutionary structure of the learning task and thestochastic and reversible dynamics of the backgammon game itself than withthe advanced learning techniques.<p><!WA14><a href="http://www.demo.cs.brandeis.edu/bkg.html"><!WA15><img src="http://www.demo.cs.brandeis.edu/bginit.gif">go to backgammon page</a><h2> Massively Parallel Genetic Programming</h2><br>The SIMD power of the Maspar MP2 is harnessed to perform geneticprogramming at top speed. See our <!WA16><Ahref="ftp://ftp.cs.brandeis.edu/pub/faculty/pollack/gp2.ps.Z">chapter</a>in Advances in GP II, forthcoming from MIT Press. <p><h2> Coevolutionary Genetic Programming</h2><!WA17><A href="http://www.demo.cs.brandeis.edu/spiral.html"><!WA18><img src="http://www.demo.cs.brandeis.edu/smspiral.gif"></a>Using a competitive fitness, we were able to evolve an elegantsolution to the intertwined spirals problem. This function uses 52 GPprimitives and broke the plane into two subproblems which added up toa spiral.</h2><p><h2> <!WA19><a href="http://www.cs.brandeis.edu/~sevan/GNARLProject/gnarl.html">Massively Parallel Neural Networks</a></h2><br>The MP2 is used for collective action bysmall neural networks, and for evolution across network architectures.<p><h2> High Capacity RAAM's</h2><br>Using an IFS concept for decoding,we find very large capacity RAAM's. Here is apicture of tree depth equivalence classes for a particulartwo dimensional attractor.<p><h2> The Mind's Eye Project</h2><br>The MP2 is also used for rapid construction of neuralnetwork IFS-like fractal attractors as a basis for memory. <b><!WA20><A href="http://www.demo.cs.brandeis.edu/bwifs.html">Here is a gallery of images...</a></b><p><h2> Evolution of Communication</h2><br>We evolve neural networks to put in agents who cooperate and competeand communicate on robotic tasks. Students are needed to work on more realisticsimulation in preparation for downloading intoactual real robots. <em> If and only ifyou use ghostscript as your ps viewer, you can watch a <!WA21><A href="http://www.demo.cs.brandeis.edu/animate.ps"> simulation</a> ofour evolved agents in action.</em><p><h2> Robot Building </h2><br>Thanks to the <!WA22><a href="http://www.nsf.gov">NSF</a>,we are building anew robotics facility at Brandeis, complete with LEGO's,small metalworking equipment, and electronics capabilities.Opportunities abound, especially for people with electronicinterfacing and machine shop experience. <br></body></head>⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?