neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (10/19/90)
Neuron Digest Thursday, 18 Oct 1990 Volume 6 : Issue 61 Today's Topics: A neural net who plays chess Re: A neural net who plays chess Re: A neural net who plays chess Re: A neural net who plays chess Re: A neural net who plays chess Re: A neural net who plays chess NN for weather forecasting Neural Net Visual Servoing Systems posting for Neuron Digest Weather forecasting References Re: Weather forecasting References Kohonen's Self Organizing Map for Pattern Recognition neural nets applications to controls REFERENCES NEEDED Seismic analysis Neuro-Nimes'90 COGNITIVE SCIENCE/HCI INITIATIVE Talk at American University Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" Use "ftp" to get old issues from hplpm.hpl.hp.com (15.255.176.205). ------------------------------------------------------------ Subject: A neural net who plays chess From: swrinde!cs.utexas.edu!sdd.hp.com!zaphod.mps.ohio-state.edu!uwm.edu!csd4.csd.uwm.edu!markh@ucsd.edu (Mark William Hopkins) Organization: University of Wisconsin-Milwaukee Date: 04 Oct 90 07:17:59 +0000 I've been sitting on this extremely simple but potentially very powerful design which would enable a neural net to learn to play chess effectively, starting even from scratch. Haven't gotten around to coding it yet, or testing it, but if you'd like to try it yourself, I've posted a specification below. (1) ARCHITECTURE The design is actually of a hybrid architecture: embodying both your typical game-playing/search AI program and a typical neural net. Simply put, the search program (the "left hemisphere") acts as a 'predictor'. It searches out the best configurations N moves ahead of the current configuration. The neural net (the "right hemisphere") acts as an 'evaluator'. It's purpose is to fill in the role formerly held by the hard-coded heuristics you usually would have otherwise found in a chess-playing program. These two components reenforce each other in such a way that the evaluator gradually learns to "squash" more and more of the N-move search tree into a static global evaluation function. This, of course, serves to increase the effectiveness of the predictor. (a) Evaluator The input to the evaluator is a board configuration which contains information to allow you to completely determine the position of any piece on the board (or whether it has been captured or not). It can be as simple as a matrix, or as complex as a net with 'diagonal detectors', 'knight-L detectors', etc. built directly into the input representation. Alternatively, it might just be a list of board locations indexed by piece. The output is a number, say, between -1 and 1, indicating the value of the board. A 1 indicates that the board is a winning board from the program's point of view, and -1 ... a losing configuration. (b) Predictor The predictor is a program to implement search, such as minimax with a cutoff. At the cutoff, the evaluator is used to estimate the configuration in question. This allows the predictor to return with a value in short finite time. It also contains the body of rules that determine what comprises a valid move, and a winning, losing, or stalemate configuration. (2) INITIALIZATION The evaluator MAY be initialized by training it using an already available evaluation function as a supervisor to incorporate already developed expertise (why re-invent the wheel?), or can simply be set to a random configuration to simulate the process of learning the game from scratch. (3) THE PLAY CYCLE (a) Move generation As mentioned above, to generate a move, the predictor performs a search and uses the evaluator at the cut-off point. It comes back with a best possible move AND an estimate as to the value of the current configuration. (b) Learning Upon completion of a move, the evaluator is updated with the training pair (value, configuration). After the opponent makes a move, the play cycle is then resumed at (1). Play continues, of course, until the program or its opponent wins (or, using standard rules, until a draw or stalemate happens). (4) EXPECTED PERFORMANCE Provided the network was designed with care, I expect to see impressive gains in playing ability over time. (a) Static evaluation and search-tree squashing. As time goes on, because of the feedback the predictor is providing to the evaluator, the evaluator will actually learn to make static evaluations of the current board, effectively squashing the N-move search tree into a single global evaluation. This kind of static evaluation in lieu of searching is precisely what good chess players claim to be able to do. (b) Spontaneous emergence of openings. After getting its butt severely kicked the first few hundred thousand games (again, we're talking about the case where the evaluator starts from zero-knowledge), natural pathways will develop that correspond to the openings typically taught to a novice. (c) Spontaneous emergence of endgames. If the input network architecture has been designed properly, the evaluator may be able enough to make generalizations (say, of tranalation-invariance) to embody knowledge of the more-or-less standard endgame configurations. (5) EMBEDDING THE PREDICTOR An interesting extension to the design specified above would be to embed the search control structure of the predictor itself in a neural net. If this is done, then the hybrid "crutch" (that's all it really is: a crutch) can finally be discarded. ------------------------------ Subject: Re: A neural net who plays chess From: dave@cogsci.indiana.edu (David Chalmers) Organization: Indiana University, Bloomington Date: 06 Oct 90 08:20:21 +0000 >the search program (the "left hemisphere") acts as a 'predictor'. >[[...]] There is the seed of a very nice idea here. i.e., the best evaluation function in chess is a fixed point of the "minimax search" operator in function space. (Spelled out, that means that for an optimal evaluation function E from positions to values, evaluating the position via minimax search, applying E at the end of the search tree, should provide exactly the same result as applying E directly to the position.) Combined with certain constraints on E, e.g. E(checkmate position) = 1, E(checkmated position) = -1, E(2 kings) = 0, this characterization should provide a very tightly constrained evaluation function that would play a mean game of chess. So, is there any way to take advantage of this by using methods of fixed-point analysis? I confess to know almost nothing about computer chess, so I don't know if the idea is commonplace there. Your suggestion of bootstrapping the evaluator with the predictor is a nice way of taking advantage of this property. There are a couple of problems, though... > (b) Learning > Upon completion of a move, the evaluator is updated with the training pair >(value, configuration). >(4) EXPECTED PERFORMANCE > Provided the network was designed with care, I expect to see impressive >gains in playing ability over time. > (a) Static evaluation and search-tree squashing. > (b) Spontaneous emergence of openings. > (c) Spontaneous emergence of endgames. You may be overestimating the capabilities of networks just a tad. Network training isn't magic... networks are ultimately pretty simple structures, not great for computing incredibly complex mappings like the ones required here (yes, I know that you can get universal approximation, but only at the cost of generalization). After all, if you expect the above to work, why not just try: train the network directly on (position, move) pairs directly derived from the games of Kasparov and other grandmasters. We'll get a grandmaster network, easy! And once we've done that, why not try: train a network to simulate the total performance of a person in all domains by training it from data provided over the lifetime of a person: (sensory input at time t, motor output at time t). Human intelligence, just like that! Of course, you'll need to use a recurrent network, but that's no problem, we can just train it up with back-propagation through time or the Williams/Zipser algorithm. I think you see the problem by now... >(2) INITIALIZATION The evaluator MAY be initialized by training it using >an already available evaluation function as a supervisor to incorporate >already developed expertise (why re-invent the wheel?), or can simply be >set to a random configuration to simulate the process of learning the >game from scratch. One thing you don't want to do is to start completely from scratch. Even bootstrapping needs to start from somewhere. All this would guarantee is that you would arrive at *some* fixed point of the minimax operator -- but there are a lot of these -- the constant zero function for instance. You have to make your initial evaluator minimally competent (incorporating at minimum the constraints mentioned in the first paragraph above, and preferably more) if you want bootstrapping to get anywhere. Actually, as things stand there's no need for the network to play actual games at all. You just generate positions randomly, calculate the minimax+evaluator prediction, and train the evaluator on that. (Only one step of minimax search is required, nothing deep.) In fact, your proposal doesn't incorporate any feedback from wins and losses, so playing games doesn't serve any purpose at all (except perhaps to confine the training of evaluation to a "reasonable" portion of position space). And by turning it into a fully-supervised learning problem, you've made it much easier than it would be with the kind of reinforcement learning that would be required from playing actual games, with all the accompanying credit-assignment problems. Nice idea, and I'm not sure that it couldn't be taken advantage of in various interesting ways. Nothing about this need be specific to neural networks, incidentally -- any kind of learning system would do. Just one that has an architecture that guarantees both complete learning and perfect generalization for the problem at hand. Which sadly doesn't exist in 1990, but hey, that's only a minor point... Dave Chalmers (dave@cogsci.indiana.edu) Concepts and Cognition, Indiana University. "It is not the least charm of a theory that it is refutable." ------------------------------ Subject: Re: A neural net who plays chess From: markh@csd4.csd.uwm.edu (Mark William Hopkins) Organization: University of Wisconsin-Milwaukee Date: 08 Oct 90 23:58:34 +0000 In article <62519@iuvax.cs.indiana.edu> (David Chalmers) writes: >... In fact, your proposal doesn't incorporate any feedback from wins >and losses, so playing games doesn't serve any purpose at all (except >perhaps to confine the training of evaluation to a "reasonable" portion >of position space)... It wasn't made explicit, but the search program (where the rules and winning, losing and draw conditions are stored) provides the feedback to the evaluator. ------------------------------ Subject: Re: A neural net who plays chess From: mahler@latcs1.oz.au (Daniel Mahler) Organization: Comp Sci, La Trobe Uni, Australia Date: 09 Oct 90 13:22:08 +0000 I did my honours thesis on using NN's learn to play games. I used a program based on the principle you describe to learn noughts and crosses. There are 2 points worth making about this aproach. 1) It is essentially an extension of the learning technique in Samuel's checkers program. Samuel used an evaluation function that was a linear combination of given feature functions. These weights were optimised by getting the minimax value for a position using the current weights, and using it as a training signal. Samuel worked before the perceptron era, so he did not call it a perceptron. His technique worked because the feature selectors were designed by a human expert. I tried the natural extension: multilevel perceptron with bp, working from a simple representation of the board and trying to see if the net can develop tactical and strategic concepts. I started with random weights, because my interests were mainly theoretical (that's what I always say when something doesn't work 8>) ) Part of the reason for my techniques failure is the second point. 2) The minimax operator does not have a unique fix-point (I like this terminology, shame it's to late for my thesis). In fact it has infinitely many, as any constant value function is a fix-point. This means, in NN terms, lots and lots of local minima. Supplying known strategic features as inputs simplifies the energy space. This kind of learning is probably affected by game tree pathology, especially in the early stages. I think the solution to this problem is to incorporate a genetic techniques with large populations and/or use reinforcement learning (Barto & Sutton's Arp element). Samuel (and I) used a sort of genetic technique with population size 2. There are 2 nets that play each other called champion and challenger, or master and apprentice. The weights in champion are fixed. Challenger learns by the method under discussion. When challenger outperforms champion by some criterion, it becomes champion and a new random challenger is generated. This leads to peculiarities which i ascribe to 'inbreeding'. The graph of the number of games taken by each successive challenger to become champion is a saw-tooth. Each new challenger takes longer than the last, (which I interpret as each champion being better) until suddenly there is a champion that lasts a very short time, and then the cycle repeats, though the peaks seem to get higher. (fractals fanatics may wonder if there are peaks within peaks). I take this to mean that eventually there evolve nets designed to beat the champion though they can be beaten by a random player. Daniel Mahler beating D A ------------------------------ Subject: Re: A neural net who plays chess From: mahler@latcs1.oz.au (Daniel Mahler) Organization: Comp Sci, La Trobe Uni, Australia Date: 10 Oct 90 08:46:46 +0000 Several people have pointed out that my claim about trivial fixpoints is not true as your program should recognise terminal game positions and give absolute value to these. You of course have to do this to give bp some goal directed bias, but for a game of any complexity there will be only few terminal nodes in your search horizon. There will usually be some in chess but NONE in go or othello until the endgame. For noughts and crosses we can overcome all the problems by tabulating all the legal positions and their true min-max values. This is feasible as there are less than 3^9 legal positions to evaluate. It happens in a few seconds of real-time. Then you do not have to worry about using genetics to avoid inbred strategies, but you do not have to worry about neural nets or learning either. I gave my tic-tac-toe program a limited horizon because I was intersted in going onto go (bravery is usually born of innocence/ignorance). Overall, I do not think this generalised Samuel's technique is viable for learning from scratch, which is what machine learning people really want. :) (For this Lenat's Eurisco, genetics or Barto&Sutton seem more promising) First it is computationally very expensive: to produce the teaching signal you must do a search and eavaluate a neural net at each terminal node. Initially this signal will be poor anyway. It is likely that tree pathology (Nau's work) becomes a significant factor under these circumstances. Also for a complex game the net must learn to cluster positions by tactical/stragic concepts from the extremely small percentage of all legal positions it will ever see. However, a slight modification of Samuel's idea seems good. Use feature detectors designed by experts as inputs to a small net. This will then be Samuels method extended to allow nonlinear combinations of these features as evaluation functions. Daniel ------------------------------ Subject: Re: A neural net who plays chess From: markh@csd4.csd.uwm.edu (Mark William Hopkins) Organization: University of Wisconsin-Milwaukee Date: 11 Oct 90 06:50:22 +0000 In article <8966@latcs1.oz.au> mahler@latcs1.oz.au (Daniel Mahler) writes: On the emergence of goal-directed behavior: > Several people have pointed out that my claim about trivial >fixpoints is not true as your program should recognise terminal game >positions and give absolute value to these. You of course have to do >this to give bp some goal directed bias, but for a game of any >complexity there will be only few terminal nodes in your search horizon. >There will usually be some in chess but NONE in go or othello until the >endgame. The evaluator is always learning based on inputs from positions N moves ahead. Initially it will only have substantial input within N moves of the end, but this learning propagates as it takes THIS input in later games and evaluates N moves from *it*: now 2N moves from the end. And so on... The evaluator's trying to converge onto a non-trivial fixed point. The goal directed behavior should propagate from bottom up, as it were. If it really learns, the first thing you'll notice is that it'll progressively avoid bad openings one by one because they lead to quick endings. Eventually, by exclusion, you'll have the good openings left. In that sense, the bottom-up learning also leads to top-down learning. I think the relevant issues are probably whether the network will be big enough to handle the knowledge a typical chess master has of the state space, and whether it can learn quick enough without having to beating it against a wall a hundred thousand times just to even get it to do something as easy as picking up a simple boolean function? On using hard-coded feature detectors to expedite learning: > However, a slight modification of Samuel's idea seems good. Use >feature detectors designed by experts as inputs to a small net. This >will then be Samuels method extended to allow nonlinear combinations of >these features as evaluation functions. What's to keep the evaluator neural net from spontaneously generating nodes that emulate feature detection in its hidden layer(s) anyhow? As a concrete example on a simpler game try this design on tic-tac-toe: LAYER 3: 1 node (output). LAYER 2: N nodes (try N = 12). LAYER 1: 9 nodes (input .. connected to tic-tac-toe grid) For input, a square occupied by the program's piece evaluates to +1, a square occupied by an opponent's piece to -1, and an empty square to 0. Every cell is connected to every other cell in an adjacent layer. Use back-propagation *with thresholds*. Take your learning input from a minimax search program that looks merely 1 move ahead using the output node of the evaluator function to resolve all non-winning/non-losing/non-draw positions. When using +1, and -1 as limmiting activation values the sigmoid function 1/(1 + e^(-X)) becomes tanh(X), by the way, and the error correction factor Y(1 - Y) becomes 1 - Y*Y. I predict two things will happen: (1) The program, when playing X, will open in the middle square with increasing frequency. When playing O, it will move into a corner in response to an X placed in the middle more and more frequently. (2) Some of the hidden-layer nodes will evolve spontaneously into 3-in-a-row detectors. ------------------------------ Subject: NN for weather forecasting From: muttiah@stable.ecn.purdue.edu (Ranjan S Muttiah) Organization: Purdue University Engineering Computer Network Date: 05 Oct 90 02:35:54 +0000 I am looking for any references that deal with this topic. Has anyone looked at time series using neural net ? email me if you would please. ------------------------------ Subject: Neural Net Visual Servoing Systems From: Keith Nicewarner <nicewarn@ral.rpi.edu> Date: Thu, 11 Oct 90 21:36:57 -0400 I am currently developing a frame-rate visual servoing system for a PUMA robot arm. I am using a camera mounted on the gripper of the robot for feedback. The objective is to use this information to guide the gripper towards an object in the field of view, matching the position and orientation of the gripper with the object. My current options are to use localized Jacobian techniques or traditional inverse-kinematics, assuming that the position and orientation of the object (specifically, a cyllinder of known diameter) can be derived using special markers on the object. I have looked at a few neural network systems which deal with the problem of visual servoing, specifically, the various hand-eye coordination systems developed by Grossberg and Bullock. These models, however, were developed with only the intent of simulating and understanding biological systems. My needs require a practical, fast, and easy to impliment system which doesn't necessarily have to be all that accurate. Does anyone have any information about neural network visual servoing systems, or more generally, hand-eye coordination or object tracking systems? Feedback will be greatly appreciated. Keith Nicewarner Center for Intelligent Robotic Systems for Space Exploration Renssalaer Polytechnic Institute ------------------------------ Subject: posting for Neuron Digest From: daft@debussy.crd.ge.com (Chris Daft) Date: Fri, 12 Oct 90 21:17:41 -0400 A while ago I asked net-people for their ideas on what the best work was on neural nets and image processing/image analysis. I got lots of helpful mail and have now written a review article on it. Some folks asked if they could get a copy of this: if you want one, send me a message. Unfortunately, it does not exist as a file you can ftp, but I will be happy to snail-mail it to you. The paper will be published in the proceedings of the 1990 IEEE Ultrasonics Symposium. Chris Daft, GE Corporate R&D Center. ------------------------------ Subject: Weather forecasting References From: muttiah@stable.ecn.purdue.edu (Ranjan S Muttiah) Organization: Purdue University Engineering Computer Network Date: 13 Oct 90 04:21:32 +0000 I wish to thank those who responded to my request. There was one paper by Widrow and Hoff for weather forecasting (using madaline ?) in the early 60's that I'm trying to locate. Anyone know of this paper ? ------------------------- For predicting time series: Moody, J. Darken, C., 1989, Fast Learning in Networks of Locally-Tuned Processing Units, Neural Computation, 1(2), pp. 281-294 For predicting weather: Rogers, D., 1990, Predicting Weather Using a Genetic Memory: A Combination of Kanerva's Sparse Distributed Memory with Holland's Genetic Algorithms, Advances in Neural Information Processing Systems, vol 2, pp. 455-464 One paper I have found particularly interesting is Predicting the future: A Connectionist Approach A.S Weigend, B.A Huberman and D.E Rumelhart, 1990 Stanford Universtity technical report Standford-PDP-90-01 Submitted to the International Journal of Neural Systems. in which the authors describe their work in perdicting future values of chaotic time series using backprop networks. It is based on a previous paper: Nonlinear Signal processing using neural networks: prediction and system modelling. A.S Lapedes and R.M Farber Technical report LA-UR-87-2662 Los Alamos National Laboratory, 1987 ------------------------------ Subject: Re: Weather forecasting References From: aboulang@bbn.com (Albert Boulanger) Organization: BBN, Cambridge MA Date: 13 Oct 90 20:01:29 +0000 One other excellent paper that represents this approach (chaotic time series prediction) is: "Nonlinear Forcasting as a Way of Distinguishing Chaos from Measurement Error in Time Series" George Sugihara & Robert M. May Nature, Vol344, 19 April 1990, 734-741 They give examples where the nonlinear perdiction technique works AND (contrary to what one sees in the AI world) does not work. Regards, Albert Boulanger aboulanger@bbn.com ------------------------------ Subject: Kohonen's Self Organizing Map for Pattern Recognition From: "M.Elif KARSLIGIL" <ELIF%TRYILDIZ.BITNET@CUNYVM.CUNY.EDU> Date: Fri, 12 Oct 90 17:35:09 -1100 I'm looking for some information about Kohonen's Self Organizing Map for Pattern Recognition. Any references will be much appreciated.. Elif Karsligil Yildiz University Computer Sciences and Engineering Dept. E-Mail:Elif at Tryildiz Address: YILDIZ UNIVERSITESI EHBAM 80150 Yildiz-ISTANBUL TURKEY ------------------------------ Subject: neural nets applications to controls From: ELEE6UG@jetson.uh.edu Date: Sat, 13 Oct 90 21:48:00 -0500 Hi I have just been introduced to the neuron digest . I am a graduate student at the University of houston.My current interest is neural networks applications to linear and nonlinear controls. I was working in nonlinear dynamic controls when I came accross a paper by CHARLES W ANDERSON -"Neuronlike Adaptive Elements that can Solve difficult Learning Control Problems." IEEE Transactions on systems Man and Cybernetics-vol-smc-13,no 5,september/ october 1983 The paper described a neuronlike adaptive element system (search and critic) to control a clasical control problem, stabilizing a cart mounted inverted pendulum system.The very idea to control a system(nonlinear in this case)without knowing the dynamics of the system got me interested in NN. However all the other papers (that I have read) ,including the one mentioned, that followed to solve a similar problem, never showed results when the system was REALLY nonlinear.for example ,for those familiar with the above mentioned problem ,results were not available when the starting angle of the inverted pendulum is large (more than 54 degrees).For a small initial angle the system can be approximated to a linear system. I am interested in hearing from people that share a comnmon interest. I have been introduced to NN only a few months ago .I would appreciate if some of you could suggest a good reading on NN and controls. Sidharth Sibal Email:ELEE6ug@uhvax1.uh.edu ------------------------------ Subject: REFERENCES NEEDED From: qian@icopen.ICO.OLIVETTI.COM (DA QUN QIAN) Date: Sun, 14 Oct 90 09:53:55 +0100 I am looking for references on automatic modifcation of structure of a neural network by learning algorithms(i.e. automatically add some links and nodes to a neural network or remove them from the neural network. Any information are welcomed. Thanks in advance. Qian Da Qun Artificial Inteligeence Center Olivetti Nuova ICO 3 Piano Via Jervis 77, 10015 Ivera(TO) ITALY Eamil: qian@icopen.ico.olivetticom ------------------------------ Subject: Seismic analysis From: quark@xroads.UUCP (Jerry Rightnour) Organization: Crossroads, Phoenix, AZ 85046 Date: 14 Oct 90 23:12:38 +0000 I saw the following presentation announcement in the NIPS 1990 schedule in the Neuron Digest V6 #58. >App7* "Seismic Event Identification Using Artificial Neural Networks", >by John L. Perry and Douglas Baumgardt. I know some geologists who might be interested in this work. Could someone provide an address (email or USmail) or phone number for contacting the authors? Also, any other references to work on computer analysis of seismic signals would be appreciated. (Not limited to neural net approaches.) Jerry Rightnour work 602-862-5918 14020 N. Black Canyon #2024 home 602-375-8409 Phoenix,AZ 85023 email quark@xroads.uucp \ / C r o s s r o a d s C o m m u n i c a t i o n s /\ (602) 941-2005 300|1200|2400 Baud 24 hrs/day / \ hplabs!hp-sdd!crash!xroads!quark ------------------------------ Subject: Neuro-Nimes'90 From: Francisco Castillo Cobo <castillo@eel.upc.es> Date: 18 Oct 90 13:28:00 +0100 I would like to take this opportunity to invite everyone in the field of Neural Networks to the upcoming Neuro-Nimes '90 conference to be held from Nov.12-16,1990 in Nimes, France. It also includes, apart from the conference, some tutorials and an exhibi t ion. For more information please contact: Marie-Martine Sainflou EC2 269-287, rue de la Garenne 92024 Nanterre Cedex - France Tel:+331.47.80.70.00 Telex 612 469 Telefax: +331.47.80.66.29 Please do not respond to this message! ------------------------------ Subject: COGNITIVE SCIENCE/HCI INITIATIVE From: Dr L S Smith (Staff) <lss@compsci.stirling.ac.uk> Date: Wed, 17 Oct 90 14:49:58 +0100 COGNITIVE SCIENCE/HCI INITIATIVE Department of Psychology University of St Andrews Scotland, UK and Centre for Cognitive and Computational Neuroscience University of Stirling Scotland, UK 2 POST-DOCTORAL RESEARCH FELLOWSHIPS investigating psychological and neurocomputational processes of visual word recognition. This project represents a major collaboration between the Department of Psychology, St Andrews (a leading centre for human perceptual research) and the CCCN, Stirling (a leading centre for neural network research) to develop a new computational model of visual word recognition. Applications are invited for the following 2 post-doctoral fellowships within the project: Post 1 (tenured for 3 years, based at Department of Psychology, St Andrews University) will involve developing fresh perspectives on the neural modelling of visual word recognition from human experimentation. The data from these experiments will form the basis for the computational modelling in the project. Applicants should have experience in human experimentation in cognitive science or perceptual research, be well acquainted with the use of computers in experimentation, and have some knowledge of neural network research. Post 2 (tenured for 2 years, based at Centre for Cognitive and Computational Neuroscience, Stirling University) will involve setting up and developing a new computational model of visual word recognition which combines the findings from St Andrews with fresh perspectives on neurocomputational processing. Applicants should have experience or interest in neural computation/connectionism and have a background in one or more of the following: computing science, psychology, mathematics, physics. Starting salary for each post will be on the 1A scale for research staff (up to \pounds 18165 pa). Both posts are scheduled to start as soon as possible in 1991. Application forms and further particulars for both posts can be obtained from The Director of Personnel Services, College Gate, St Andrews University, St Andrews, Fife, KY16 9AJ, to whom completed applications forms together with a CV should be submitted to arrive no later than November 30th 1990. Further information can be obtained informally from: (Post 1) Dr Tim Jordan at St Andrews (tel.0334 76161, ext 7234) (Post 2) Dr Leslie Smith at Stirling (tel.0786 67435, direct line) Previous applicants for these posts need not re-apply. Both Universities operate an Equal Opportunities Policy. ------------------------------ Subject: Talk at American University From: masud cader <CADER%AUVM.BITNET@CUNYVM.CUNY.EDU> Date: Fri, 12 Oct 90 02:15:55 -0400 ANNOUNCEMENT: Seminar Series Department of Computer Science and Information Systems The American University Paul J. Werbos National Science Foundation Neuro-control and Fuzzy Logic: Connections and Designs 2:00 pm Monday, October 29, 1990 LOCATION: To Be Announced Astract: This presentation will focus on application of artificial neural networks to control tasks. Concentration will be on contrasting Fuzzy Logic with the interpolation power of current neural architectures, as well as integrating fuzzy logic and neural systems. This discussion will take place in the context of control theory. This talk can be considered as a follow-up to Dr. Werbos' talk at IJCNN-WASH-90. DIRECTIONS: To get to American University (Washington DC) by public transportation, there are many alternatives: 1. If you are coming from Downtown (DC) you can take the bus N2, or N6 (Friendship Heights direction) at Dupont Circle. 2. If you are taking the metro (underground), please take the red line to Tenleytown. From there, you can take M4 city bus or the AU shuttle bus, or walk along nebraska avenue (about 15 min to walk to AU). 3. You can also stop at Friendship Heights, and then take the N2, N4, or N6. If you are driving a car, please park at the METROPOLITAN MEMORIAL UNITED METHODIST CHURCH. American University is located at the intersection of Massachusetts Ave (NW), and Nebraska Ave. The Department of Computer Science and Information Systems is located in Clark Hall, 4400 Massachusetts Ave., NW, Washington, DC. 20016. Call 202/885-1470 for further information, or send e-mail. ------------------------------ End of Neuron Digest [Volume 6 Issue 61] ****************************************