Junior Students
Computational Science
Photo Album
The Chais Teachers Centre
Computational Science
  Computational Science Project 
 Rational and Syllabus    The computational Science Project 
 Final Projects Abstracts 2008   The 2014 Schwartz-Reisman Competition for Scientific Excellence 


Final Projects Abstracts

 The Computational Science Program


The Following Abstracts describe some of the final projects of the 4th finishing class of Computational Science (CS) at Hemda. The ideas for all the projects were chosen by the young students themselves and cover a wide variety of issues from different disciplines.


All projects were developed with Matlab software and employ Neural Networks and Genetic Algorithms - the main learning chapters of the 3rd year of the program (12th grade). These Algorithms are based on neural computation principals in the brain, Parallel Distributed Processing and biological evolution, and can be implemented in various computational tasks: face recognition, hurricane course prediction or autonomous robotic navigation.


Insulin Dosage

Galy Milman Gymnasia Herzliya High School

Merav Pnkas Ironi Yud Dalet High School

Instructor: Dr. Eyal Cohen


The software developed in this project allows Type-1 diabetics using an  insulin pump to determine the exact amount of insulin to be injected. The program scans the patient's past tests results and learns his/hers reactions to insulin under a range of conditions. At the end of the process, the program is capable of determining the amount of insulin to be injected relative to a target blood sugar level. The required amount of insulin is determined by the present blood-sugar level, physical activity expected and amount of carbohydrates the patient plans to consume in a given period of time.

In order to carry out the project, Galy collected data from 256 self-tests and fed them into a neural-network which was trained with test data. Galy and Merav carried out thorough tests in order to determine the optimal parameters for the neural-network's architecture and dynamics.

In the final, clinical test, the system succeeded in improving the level of patient accuracy from 27% (with the manufacturer's methods) to 70%! In the graph below, the correct values are marked in red and the computerized values in blue. 

Another advantage of the developed system, which employs a learning algorithm, is its ability to adapt to the personal needs of each user and update its computations in accordance with changes occurring over the course of the year and changes in the physiological condition of the diabetic (essential for Type-1 diabetics due to their rapid growth during adolescence). 


Identifying Internet Servers

Ido Greenberg and Yonatan Tzederbaum Alliance High School

Instructor: Dr. Eyal Cohen


Ido and Yonaton's project deals with identifying the manufacturer of HTTP servers on the internet. Such identification is of major importance in the field of network security. The process of server identification is based on the site's response to various requests sent by the user's ("client") computer.

Response samples were collected from servers located at the Check Point Technologies lab. The company engineers cooperated in the initial processing of the data. Responses were collected from 249 different servers. Responses text lines were then converted into numerical codes that included information extracted from the server's responses. These data included: the Header type (from among 17 possible types), location, number of characters and first letters in the error announcement, and the error code. These data served as the Matlab-adapted database employed in the training of the neural networks. The process required a unique network architecture which was developed by the students, and operated in separate modules (see illustration bellow).


The neural networks were trained to distinguish between Apache servers, IIS servers and "others". The training process was carried out on 174 servers and tested on another 75. The mean success rate was remarkable: 98%. Furthermore, the final version succeeded in identifying an Apache server which disguised itself as a Microsoft server. 


Autonomic Control Systems for Computerized Football

Alon Naor, Tomer Nachum Ironi Yud Dalet High School

Alon Reshelbach Lady Davis High School

Instructor: Dr. Eyal Cohen


The title of this project may be misleading. Its goal is not a football game, but the following interesting research question, about parallel processing systems designed to carry out a number of tasks (for example, the brain): Is a single, multi-task system preferable to a complex system made up of separate modules for each task?


In order to examine this issue a simulated football game (see figure below), in which two teams of two players each, were programmed. Each virtual player processed game input-data which included (1) the direction of his teammate, (2) the direction of the ball or the rival team's goal, (3) the rival players' directions.

Each player was managed by a neural network (whether simple or complex) relying on the game input-data to determine whether to kick the ball, in which direction and where it should advance on the next step.


The neural networks acquired the knowledge of how to manage the players by means of complex Genetic Algorithms, based on Darwinian evolution laws adopted for computer programs. The evolutionary process included, for each generation of players, the creation of mutants (with one parent for each offspring) and reproduction re-combinations (with a pair of parents). An additional interesting point in the research is that the decision of who will pass on its "genes" to the next generation (survival of the fittest) was determined by the group's achievements and not on the achievements of the individual.

Intensive research work, which included simultaneous runs on ten computers, revealed that a complex brain was superior in achievements in all the parameters examined: general score, ability to score goals, defending the goal, catching the ball, cooperation between the group's players and movements. These results lean toward the advantage of a complex modular architecture in planning multi-task systems.

Another lesson learned, and no less important, is that it is preferable to begin with a focused and defined research proposal in order to solve complex questions.


Hurricane Path Forecasting

Eli Dayan Ironi Dalet High School

Yizhak Hefner Alliance High School

Instructor: Dr. Eyal Cohen


Eli and Yitzhak designed software which predicts, with great accuracy, the location of a hurricane's center 6 hours in advance (see illustration bellow, in which the hurricane's path is marked in blue and the software prediction in red).


The prediction algorithm relies on hurricane data that includes the date and time, geographic location, speed and direction of progress, barometric pressure and winds speed. The data was downloaded from the Atlantic Oceanographic and Meteorological Laboratory site which covers hurricanes in the Atlantic Ocean between the years 1851 2007. Due to limitations in computing power the prediction was based on a small sample of the available database. Hurricanes parameters were scanned by means of "trained by

examples" neural network.. At the end of the scanning process the network accurately predicted new hurricane paths those not included in the already scanned database.

Prediction accuracy was achieved due to the use of a smart coding scheme of the geographical data. In comparison, one can see the American Hurricane Center's rate of accuracy in the next illustration. The geographic scale in both illustrations is equal, with the difference in time between the hurricane's initial location (orange dot) and the first prediction (in black) is 10 hours.


The software developed is generic and may be employed throughout the world in all areas in which there is a hurricane database.


Biometric Iris Recognition

Liran Hasson Lady Davis High School

Instructor: Dr. Eyal Cohen


The software developed by Liran offers identification of people based on images of their irises ("biometric iris recognition"). This software is innovative in its use of the image's raw data and not data initially processed to create a complex "iris code". In search for a database, Liran contacted the University of Bath in England and downloaded 1,000 photographs of the eyes of 20 people from its database.


The photographs were converted into the Matlab format. Then, Liran developed a computerized-vision algorithm for fast identification (less than a second) of the pupil's location.


Three circles were then cut from the image of the iris (see red circle in the picture above) to be used as the raw input-data for the identification. These data were fed into artificial neural networks trained to perform the identification task. The network examined the new data by cross-checking the results from both eyes and succeeded in correctly identifying 84% of the cases. These are excellent results considering the minute amount of image's data (approximately 0.03%) used. This is most significant when working with very large databases.



Character Recognition in Computer Games

Udi Nechmad

Instructor: Dr. Eyal Cohen


The computer-vision software developed by Udi is capable of identifying a defined "enemy" character on the screen of a computer game, for example, the left character in the picture on the right:



The software defines the character by collecting a random sample of small patches from the image on the screen. Each patch is attributed a number value 0 if it does not belong to the defined character and 1 if it is a patch taken from that character. Artificial neural networks are trained to perform the classification of the patches according to their number value (1 or 0, namely, enemy or not).  They scan all the patches and learn to attribute a number value to each patch.


In order to identify an enemy character in a new display the software first randomly samples a number a patches from it.  Each patch is then classified by the trained neural networks. If the classification result is "0" the patch is colored in black and placed back in the image.  Similarly, patches classified as "1" are colored in white. The result is that the enemy character appears in white on a black background (see the right side of the illustration below).




This type of program is useful in computer games but of greater importance in military developments of "mechanical warriors".


Nuclear Decay Factors

Roy Ben Baruch, Ziv Moskuna and Dor Cohen Ironi Yud Dalet High School

Instructor: Dr. Eyal Cohen


The goal of the project was to investigate which nuclear parameters influence the nuclear stability of a given isotope. This question is very important as there is no successful model of the structure of the atom's nucleus. It is for this reason that the pupils used the characteristics of 350 different isotopes taken from a Tel Aviv University database. An array of 100 neural-networks were trained to determine isotopic stability relying on ten parameters including: atomic mass, number of nucleons, number of protons, number of neutrons, nuclear binding energy, quantum statistics, parity, and the nuclear width and spin. The mean error of the networks performance in identifying stability was 15%. The error percentage was then re-examined to determine the isotope's stability. This time the networks' training process was carried out once again using various subsets of the ten-parameter system.


The results suggest that nuclear parameters relating to the composition of the nucleus (for example, the number of protons) are more significant in determining its stability compared with its quantum characteristics (for example, the nuclear spin). Furthermore, it was found that the most significant  parameter was the nuclear binding energy, which was enough to employ two additional parameters of the nuclear components to lessen the error rate to 10% (better than the entire parameter system!).


Performance error in percentage (height of columns) for 42 different combinations of the 10 parameter system:




bgr_spacetbot.gif (59 bytes)