How to write an academic paper?

 There are many different subproblems available in scholarly writing. The pipeline contains of:

1. article writing itself

2. converting the article into the pdf format

3. publishing the pdf file

The following blog posts investigates only the first problem and assumes that the remaining one are easy to solve. The article writing itself is the result of taking notes.. Taking notes are done with an outliner software. An outliner works quite different from normal desktop publishing software.

Before a text can be written some keypoints are needed. The keypoints are creating during the reading of existing literature. If the keypoints are available in the outline editor it is possible to write down the same information as a prose text. Prose text is equal to normal sentences which can be read by the audience. The interesting fact is, that outline editors allows to write down full paragraphs. It is some sort of raw manuscript. These sentences are full of spelling mistakes and grammar errors. In most cases they are hard to read. but they are providing a first draft edition of a text.

In the next step the draft version is copied into a new file which is mostly a text processing software like LaTeX or MS-Word. Here is the idea to rewrite the sentences and make them easier to read. Also figures are added to the text and the bilographic references are inserted. From a formal perspective, a text is a reformulated version of the notes. The user has learnted something while reading the text from other. Has written down the information in keypoints and then he has reformulated the content with his own words.

The reason why this is done by authors is because it helps to understand a subject. Most texts are written not by experts but by newbies how have discovered a topic the first time. It is some sort of paradox, that especially authors without any background knowledge are motivated to explain the subject to a larger audience. The reason is, that with this precondition the author is motivated by himself. That means, it makes sense for him to investigate a topic and write something about it.

In theory it is possible to copy and paste the keypointes which were created during the research for a topic. But most authors are trying to avoid it. Because they are assuming that self-created keypoints are containing contradicting information and out of context knowledge. The more elegant way is to rewrite the content from scratch in prose text.

From printed journals to electronic ones

 The first electronic journals were started in the 1980s.[1] In contrast to today's perspective which is dominated by detail problems like Open Access in the 1980s the debate was held in a more general perspective. It was compared what the difference is between electronic journals and printed journals. The perhaps most interesting point is, that peer review and the existence of an editor are typical elements of a printed journal. They are producing sense and it is not possible to avoid them.

Let us take a look into the workflow. According to the description an author writes a paper, submits it to the editor and the editor submits it to the peer reviewer.[1] Why this workflow is remarkable is, because it was mentioned in the year 1982 as a role model for printed journals. At this time, electronic journals weren't available or they were started slowly. And if the peer reviewer has together with the editor and the author revised the document it gets printed and delivered to the reader.

Now it is possible to explain why this workflow was used by printed journals. Suppose an author submits a manuscript directly to the printer. the result is, that many authors will do so. Lots of draft articles get printed and the amount of article would be very high. The editor and the reviewer are useful to slowdown the process. Their main task is to stop an author, select only the best manuscripts and make sure that the publication is postponed.

This kind of moderation is a reaction to the printing press. A printing press is a bottleneck. Printing a document is a costly process and it will take time It is not allowed to print a manuscript if it contains of spelling mistakes or if the quality is low.

Suppose the idea is to start a printed journal from scratch. The chance is high that this fictional journal will work with the same principle in mind like the journals from the past. That means, the author is slowdowned by the editor and the peer reviewer with the attempt that only high quality knowledge is send to the printer.

In electronic journals the printing press is missing. It is possible to publish low quality manuscripts full of spelling mistakes. The reason is, that printing out something doesn't cost anything but publication is nothing else than creating an electronic file. Do we need an editor and peer reviewer to slow down the process? It is a rhetorical question. The only reason why electronic journals have a peer review is because they are trying to emulate the printed journal system from the past. That means, that the journal gets printed and the electronic version is an addon.

Or let me explain the situation from a different perspective. It is pretty easy to explain the peer review system for printed journals. Peer review provides an important filter in the process and without peer review, a printed journal can't be created. The problem is, that it is much harder or impossible to explain the peer review process in case of electronic only journals. The basic question is, why it is not possible to publish the raw version of a manuscript?

___Peer review___

The underlying assumption for a peer review is that it helps to filter the information. this makes the peer review process a powerful instance. It is in the hand of the peer reviewer to decide if a certain manuscript gets printed. What is ignored is, that not the peer reviewer is the important instance but the printing press itself is the reason why a manuscript gets rejected. A printing press is a machine which produces costs. If a certain manuscript should be published in a journal, it has to be printed first. If the journal has a circulation of 10k copies, the printing press will print out the manuscript 10k times. This process takes time and will produce costs.

Because of this reason a printing press is used seldom. That means, it is not allowed that anybody can print out his ideas but the machine has a priority queue. This queue is not something which is determined by a journal but it has to do with the limits of the mechanical machine. That means, printing works in a sense that it will produces costs and costs need to be managed.

Let me explain it the other way around. What all the printed journals have in common is, that the printed text are spelling free and that they were peer reviewed at least 10 tiimes. Only if a manuscript is absolute perfect, and only if the journal editor is 1000% sure that the reader will need the information, he will send the manuscript to the printer.

___Technology___

From a computer perspective, electronic publication has been solved. The first programs for desktop publishing were invented in the late 1980s and since the advent of HTML and PDF it is possible to publish a document in the internet. What is missing is the cultural shift towards electronic knowledge. This shift is possible only in theory but not realized yet. The problem with electronic publication is, that the amount of information will become much higher than before. It can be compared with inventing the printing press a second time.The original printing press has increased the amount of books from a few hundreds to millions. And electronic publishing will increase the amount of books from millions to ...

___Literature___

[1] Turoff, Murray, and Starr Roxanne Hiltz. "The electronic journal: A progress report." Journal of the American Society for Information Science 33.4 (1982): 195-202.

Components for a robot control system

 The main problem in robotics is, that it is hard to define some recipes which are working well for all domains. Suppose the idea is to contruct a robot which can drive on a motorcyle, or a robot forklift which can load a cargo. What is the basic principle to control these different kind of robots?

A possible walk through to solve these difficult domains is a combination of voice control command processing, model predictive control and learned cost function. Let us go into the details. The idea behind voice control is, that the robot is controlled manual but not with a joystick but with natural language. For example, the human operator can say “robot start”, or “robot load the cargo”. This sort of interaction is important because it allows to show the entire picture of a robot domain which includes the actions not automated yet. An interaction between human operator and robot is needed, if certain parts of the control system are missing. In such a case the robot is controlled with teleoperation.

The second element of a robot control system is the mentioned model predictive control tool. MPC means to predict future system states and determine the optimal action. The last strategy on the list is a cost function. A cost function helps to guide the search in the problem space. Learning a cost function is equal to inverse reinforcement learning, which is sometimes called learning from demonstration . The idea is that during the demonstration the parameters are found which are defining which sort of behavior is wanted.

Early examples for Desktop publishing

 On the first look, Desktop publishing seems to be something realized in the 1980s and today the technology is simply there. A closer look will show, that desktop publishing was at all times an advanced technology and creating the underlying hardware and software was very difficult.

Most technology historians see the start of DTP in the mid 1980s with the Apple Macintosh computer. But at this time, the technology was used in the reality. It took some years until the first user have discovered the new options. A more practical example for Desktop publishing in the reality can be seen on the Atari ST computer. Around the year 1990 the Signum II software was available. Signum II was a graphical textprocessing software which looks similar to a modern Word software. The disadvantage of Signum was, that the software needs a lot of RAM plus an externa harddrive, so the user has to spend extra money to upgrade the Computer.

In theory, it was possible in the year 1990 to write a longer text on the Atari ST with the Signum software. Books from this time are available which are describing the workflow. But it should be mentioned, that even the Atari ST was perceived as a cheap computer compared to the Apple computers, it was some sort of advanced technology to use in the year 1990s a PC or a homecomputer to create a document. Only some enthusiast tech-pionieers have done so, but not the majority of students at the university.

So we can assume, that in the year 1990s desktop publishing wasn't invented yet. The problem was, that the computer hardware at this time was missing of larger amount of RAM and most computers had no harddrive.

From a more realistic perspective, the DTP revolution was started with IBM compatible PCs and the Windows operating system around the year 1995. At this moment, the average PC was equipped with a harddrive and large amount of memory and was able to run graphical operating systems. The average student since the year 1995 was able to type in a text on a PC. On the other hand, the sad situation is, that until the year 1995 desktop publishing for the masses wasn't there. All the books, journals and documents were created somewhere else but not on a home computer with DTP software.

This might be a be surprise, because it opens up the question how academic journals and dissertations were produced from 1900 until 1995? Like i mentioned before the workflow wasn't realized with desktop publishing. But it was working in a more distributed fashion. The interesting situation is, that all the elements of a modern text processing software were available before the 1995 but not in a single location but in different larger machines and located in different companies. For example in the early 1980s printing machines were widespread available. Not on a desktop of a single user, but in a printing house. Also high quality photography was available and the ability to create longer texts. To create a book or an academic journal before the year 1995 the workflow can be described as:

- phototypesetting

- printing machine

- entering text into a terminal at a larger mainframe

- graphic design in a dedicated company

Desktop publishing wasn't inventing book printing from scratch, but desktop publishing has combined all the steps into a single software. The difference is, that before the year 1995 academic publishing was equal to group working. The steps in the workflow have to be coordinated. In contrast, desktop publishing since the year 1995 was grouped around a single person.

___Tutorials for Academic publishing___

The untrained user may wonder why universities have no courses in which the students learn how to publish a paper. Also the topic academic publishing isn't described very well in the literature. The simple reason is that in the past it was technical not possible that somebody can write or even publish a paper. Let us assume that the normal student is living in the year 1985. At this time DTP wasn't available in the reality. Without a harddrive and high resolution graphics it is simply not possible to create an academic paper. That means, if a single student at this time was motivated to create a paper and publish it somewhere it wasn't possible.

The ability for doing so was invented much later. The first tutorials can only be written and read if the underlying technology is available. That means, if it is possible to run the MS Word software under a graphical operating system it is possible to write a tutorial how to do so. This was only possible after the year 1995.

But if desktop publishing wasn't available before the year 1995, how was it possible to fill the university library with content? Somebody needs to know how to write books, conference proceedings and papers which is published. Yes, such a meta knowledge is available but book publishing before the invention of desktop publishing is working a bit different. Academic publishing before the year 1995 was equal to group working in which larger amount of people have to coordinate each other so that at the end a printed book and a printed journal is available. It is hard or even impossible to describe the overall process in a single tutorial because each subject is handled by specialists. This makes it hard to give general advice how create a paper or how to start a new academic journal. The only thing what is available for all the book publishing is, that it will need large amount of ressources. A machine which is able to print high quality journals will cost millions of US-Dollar, and running a text processing on a mainframe will cost even more.

Creating computer games alone?

 Most computer games in the 1980s and early 1990s were created in a group work. Especially the commercial high quality titles were created never by a single author but as a combination of graphics experts, musicians, programmers and marketing experts. What remains open is the reason why? Why was a group needed, why isn't it possible to create by a single person?

The answer isn't located outside the creation pipeline, it has to do with technical restrictions. The if computer hardware is slow and if support tools are missing the only way in handling the constraints is team work. Let me give an example, The Commodore 64 had a main memory of 64 kb of RAM. The only way in creating computer games was the assembly language unfurtunately, writing such programs needs a lot of time. If someone is doing so for 6 months, he has no time for creating the graphics and the sound as well. The only way to program the code and create the graphics is, that 2 and more people are working together.

Let us imagine a different sort of technology. Today the average PC has gigabyte of RAM and lots of tools are available. in addition, high level programming language like Python are available. There is no need to write the code in assembly language. Today it is possible that a single person writes down the source code, creates the graphics, invents the game design, provides the sound effects and uploads the resulting video to youtube. if the game is smaller one, a single person can do so in a weekend. There is no need for group working because better technology is available.

But what is about the quality? Would can we expect from a computer game programmed in python by a single developer? The interesting point is, that such a game has the same quality like the games in the 1980s. The only difference is, that less ressources are needed to create such a project.

Let me explain the situation from the other point of view. Group working is an indicator that a lot of resources are needed which can't be provided by a single person. Group working means, that a project will need so much time, and so much different knowledge that a larger amount of people has cooperate to create something. Group working is needed if technology is missing. this can be shown for many different domains like video games, book publishing, construction working or automotive assembly.

Perhaps it makes sense to explain who group working can be introduced to a modern video game. Suppose the idea is to create a mini game in python which has around 500 lines of code plus some low quality sprites. It is a normal jump'n'run game without any extras. Does it make sense to see this as a large scale group working project? It is a rhetorical question, because it is a typical one man programming effort. It is nothing else but a hello world demonstration of somebody who has discovered the pygame engine and likes to learn how to use it for creating a game. It is not possible and it doesn't make much sense to ask a group of people if they want to join this project. It would take more time to contact all the people than writing the small amount of codelines alone.

The typical situation in the now is, that a single person creates such a game, uploads the video into the internet, and then will read through the comments to get feedback from the end user. Such an interaction is not equal to group working but it is something else. Group working in the classical sense means, that before something was published lots of people have to interact.

Let me give a counter example. Suppose, the python language wasn't invented yet. The only computer available is a VIC20 which has to be programmed in assembly language. Under such a constraint the same project (a jump'n' run game) is much harder to realize. It is not possible for a single person in doing so. So the alternative is, either to work in a group or not creating the game at all.

Can the success of Open Access explained only with technical reasons?

 The discussion about Open Access has become mainstream within the science community. There are dedicated conferences available and many books and papers are focussing on the problem of how to open up science. Most discussions are about the question how to convince somebody to publish a paper under open access license or if this failed, what the concrete reasons are. Describing the situation from an ideology standpoint has become the standard and Open Access is often seen as a movement similar to the Open Source idea.

To make the situation more pleasent it would help to create a working thesis. Open Access is the not the result of individual decisions but the origins are located in technical development, namely desktop publishing, full text databases and bibliographic managers. The interesting situation is, that these tools are a new development and it can be traced back, that before these tools were available nobody was talking about electronic publishing.

Perhaps it makes sense to start the journey to open access with the desktop publishing software. Suppose somebosy has installed the MS Word or LaTeX software on his computer and tries to investigate the newly discovered features. What he will recognize fast is, that he can create academic papers and expert them into the pdf format. That means, LaTeX and MS Word was created with such an objective in mind, and it is very easy in doing so. If the user is unsure he can formulate it as a question and he will receive for sure an answer in an online forum, in which other users are explaining in detail how to expert a text file into a pdf document.

The interesting situation is, that the self created pdf file is the core element of open access publication. Open access assumes, that the manuscript is available in a digital format and of course in a standard format like pdf or postscript. Open Access won't work, if somebody has written the manuscript with a mechanical typewriter, because such document can't be uploaded into the internet.

To understand the upraising of the open access movement we have to investigate the Arxiv project in detail. The server was started in the year 1991 and the only allowed content was about mathematics, physics and computer science. In the 1990s, arxiv was the only preprint server in the internet. And the reason was, that in the humanities like literature or philosophy, computer technology was not available at this time. In the early 1990s, the Personal computer was an unusual device. It was available but it was a costly device. A typical desktop PC was running with MS-DOS and the software in that time wasn't able to expert text documents into the postscript format. On the other hand, PDF wasn't invented yet.

Open Access has become a mainstream topic in the 2000s. During that period, normal PCs were able to expert documents into the pdf format with a single mouseclick and the PC was widespread avaialble. It is not surprising, that since the 2000s the amount of published electronic documents was higher than before. Apart from mathematical papers also documents with a humanities background were published in the internet.

The new development was, that former book publishing companies were no longer needed. The combination of the internet plus a desktop publishing software allows a single person to write and publish a paper and he won't need classical libraries or linotype machines.

The revolution was, that with the book printing industry every thing remains the same. A classical academic pubiishing house works the same like 20 years ago. That means, somebody sends a manuscript to the publisher, it gets formatted in electronic format and then the book is distributed to an academic libraries. The new thing is, that nobody needs this workflow anymore. Today's situation is, that the individual can decide if he likes to publish a paper in the classical way, or if he likes to simply upload the pdf file into the internet. The simple reasons why so many electornic documents are available is because it is so easy to create them.

To understand the revolution in detail we have to focus on the key component of Open Access which is a desktop publishing software. Desktop publishing means, that the former workflow was simplified. Instead of using a book publishing house, a printing company and somebody who formats a documents, the author of a paper is charge of the entire process. Desktop publishing means basically that a single person authors the mansuscript, formats the layout, creates the images, checks the bibliographic references and experts the document as a pdf file. There is no need to send the manuscript back and forth between different stakeholders, but the document is created with a stand alone pc and a pwoerful textprocessing software.

Before the advent of desktop publishing a book was created by a team. The workflow can be traced back by analyzing the book cover. In most cases it was labeled by different stakeholders. A typical book in the 1970s was equipped with an imprint of a library. That was the physical place in which the book was located. Another imprint was made by the printing house. That was a company how created the physical book. Then, an imprint was avaialble by the translator. That means, a book was translated from one language into another language. Another imprint was available from the publishing house. That was a company how formatted a document and so on.

Basically spoken, the workflow until a book was created in the 1970s was distributed over many steps. Somebody may argue, that this complicated pipeline is equal to high quality book publishing, but the more likely reason why the workflow was so complicated was is because the technology in the 1970s was low. The desktop computer wasn't invented, and it was complicated to prepare a mansucript and print it out. The internet was missing, so a library was the only way to distribute the information to the reader.

The outdated book publication process in the 1970s, and the more recent publication workflow since the 2000s are both the result of a certain technology. If a book gets created with mechanical typewriter, linotype printing machines and printed libraries, a certain workflow is needed until a book is available. And if the book was made with LaTeX, pdf files and webservers a different workflow is needed until the book is published.

The illusion of Open Science

 The dominant theory about the upraising of Open Science puts the society or governmental decision into the role of decision maker and the goal is to convince more scientists that Open Science is a great idea. So it is some sort of ideology and political movement to open up research and make the publications free to read to the world. The main reason why this plot is repeated is because it give to the stakeholders the illusion that they have everything under control. Which means, that a certain scientists can decide by himself or the political decision makers can give Open Access a higher priority or not.

A closer look onto the problem of Open Access will show, that it is not the result of a decision making process but the reason is located in technological development. The first open access server in the world was the arxiv repository. The simple reason why the project is up and running is because:

- the costs for a webserver a low in the internet age

- the existence of desktop computers and the LaTeX textprocessor is common for scientists

- if Arxiv hosts many thousands paper it is pretty easy to write more of the same content

These technical conditions have resulted into a successful prepreint server. Bascially spoken, the simple reason why Open Access was started is because the manuscripts are in the digital format available so it is easier to upload the document to a webserver.

Like all technological innovation like the car, the computer or the telephone the process was not managed and there was no higher instance who has to decided to introduce it, but if new techhology was working fine, it was adapted as quickly as possible.

It is interesting to observe the Open Access was started around the same time like the desktop computer. The first electronic scientific journals on CD-ROM were available at the same time the CD-ROM was invented, and desktop computer were there to create such content. This sounds a bit trivial but between the invention of desktop publishing and electronic publishing there is a causal relationship available.

To understand the Open Access movement we have to identify technology which supports the creation of academic papers. Potential key components are full text search engines, desktop publishing software, bibliographic databases, document formats like the Postscript and DVI standard, Unix workstations and a larger amount of people who have access to these thing. The resulting open access movement is some sort of logical consequence.

The situation has much in common with the interaction between a human and other tools, for example a hammer. If somebody has bought a hammer, he will search for situations in which he can use it. And if somebody has installed the LaTeX package on his workstaion, he will write a simple hello world academic paper next.

Is C the optimal language for programming the C64?

 The good news is, that the problem of identifying the ideal programming language can be reduced to only two: Assembly vs. C. It is pretty easy to show, that C code will run slower and needs more RAM than the manual created assembly code. A typical example would be a prime number generator.

The more interesting question is, how much exactly C code will run slower. Somewhere in the internet it was written that without using cc65 optimization techniques the C code will run 5 times slower. On the first look this speaks for replacing C code with assembly. But a closer look will show, that slowdown by the factor 5 is not very much. Suppose there is an algorithm which needs 60 seconds. If the same algorithm was rewritten in Assembly it will need only 12 seconds. In both cases, the algorithm won't run in realtime but the user has to wait in front of the machine.

The main concern with assembly language is, that the sourcecode isn't readable. Even if the programmer is using lots of comments, the code will look messy. This is perhaps one of the reasons, why C has replaced former assembly coding.

Suppose the hardware is a bit faster than the original 6502, suppose some optimizatinon techniqfues in the compiler were activated, then the chance is high, that the C code will be only 2 times slower than the assembly language counter part. This slow down is acceptable because the code is way much easier to read.

Let us make the situation more practical. A naive prime number generator works with two nested loops. Such an algorithm will run in assembly around 5 times faster then the C counterpart. But, a more advanced sieve algorithm will run much faster, no matter in which programming language it was implemented. That means, the sieve prime number generator written in C will outperform the nested loop algorithm written in Assembly language easily.

The problem with coding assembly language is, that it is difficult to write longer and more complex algorithm. Implementing a sieve algorithm in Assembly is an advanced project. That means, even an experience programmer will need many hours in doing so.

It is a known fact that during the 1980s the c language was not very popular for the Commodore 64. All of the demos were written in pure assembly code. This fact is known, because the sourcecode of the demos is available and it is normal 6502 assembly code. But what if the demo competation has the contraints, that the code has to be written in C?

since a while the C64 community has discovered the cc65 cross compiler again. They are trying to use this environment to write games from scratch. The resulting games are look not impressive as the Assembly demos, but they are easier to read and it takes less effort to code them.

A list of some games written in cc65 are available at https://wiki.cc65.org/doku.php?id=cc65:cc65-apps The quality of these games is low. They are looking like the early c64 games from the 1984 year.

An interesting side question which remains unanswered in this blog post is, if Forth can outperform the C language on the C-64. the problem with Forth is, that most existing Forth systems are only interpreters, they are not converting the code into assembly instructions so the resulting program will run slow. What is known from the MS-DOS PC is, that compiled Forth code can reach the same speed like compiled c code.[1] That means, compiled forth code will run slower than hand coded assembly code.

[1] Almy, Thomas. "Compiling Forth for performance." Journal of Forth Application and Research 4.3 (1986): 379-388.

Comparing Google scholar with Microsoft academic

 

Google Scholar was the first attempt in building a search engine for academic papers. It is mostly known for it's powerful capabilities to search inside the documents and in contrast to websites of dedicated publishers it can search in the content of all the information available. After typing in some keywords, Google scholar returns a list of useful documents, similar to the normal google web search. And very important, it provides the formatted bibtex information as well so the found knowledge piece can be cited easily.

Most users have no need to try out a potential alternative. Microsoft Academic was developed later but it is used seldom. A closer look will show, that Microsoft Academic can't replace google but it can provide useful additional information. The strength is, that that Microsoft Academic has more in common with a classical bibliographic database. All the information is labeled with tags like “operating system” or “programming language”. In addition, the user can restrict the result list to papers, conference proceedings or books. The most interesting feature of Microsoft academic is, to change the sorting order. For every paper or book the amount of raw citation is shown. This allows to answer a question like “which books for a certain topic are popular?”, or “how many papers in a subject remain uncited?”.

Let us make a simple experiment. We are restricting the year from 1990 to 2000, select the domain “programming language” and show only books ordered by the raw citation count. The result page is some sort of recommended reading list. If somebody has no experience he can read the most referenced books and its for sure, that he will like them all:

1. Edmund M. Clarke: Model Checking

2. James Rumbaugh: Object-Oriented Modeling and Design

3. Martin Fowler: Refactoring: Improving the Design of Existing Code

4. Ed Anderson: Lapack Users' Guide

In contrast to a human guided recommendation list, these four books were determined by statistical information of the citation count. They are an objective list of books who are relevant for a discipline. A good library will provide these books, because they are requested frequently.

Let us imagine the situation from a broader perspective. Suppose there is a university library which has room for only 2000 books but not more. Which books should the library buy? The answer is, that they need to buy the most requested books, because the chance is high, that the student will find what they need.

Thanks to the Microsoft academic search engine, the concrete list of books can be created easily. All what the librarian has to do is to write down, the most cited books in each subject. In the example, 500 different subjects are there and each slot is filled with the most 4 important books. The result is a small but beautiful classical printed library which will be liked by its students very much.

The most referenced books for programming languages was mentioned before. Does it make sense to reduce all the books to only 4? Yes it is possible, because the requirement is, that the imaginary library doesn't have more space. Each single subject can only be filled with a restricted amount of books and not more. If “programming language” is supported with 20 books, the other categories available will be ignored. The hard to answer question is, which four books are important enough to become a role model for a single domain?

The entire domain of programming language has around 300k publications which are papers, conference proceedings, books, and postscript files in different repositories. The reason is, that a lot of people has contributed to the subject in the past. If a library likes to store all the information, it has to become a huge building.

Coach based artificial intelligence

 

The problem why Artificial INtelligence is difficult to realize is not located within the machine itself. Because software engineers are experts in writing software. The sourcecode compiles into binary code and if the compiler doesn't find a mistake, everything is working fine. The more demanding problem is, if a software is working fine on the first look but at the same time it is doing not what is expected. That means, the program compiles great into binary code but the robot isn't able to solve a task.

Fixing such problems is much harder and if it is not done another failed robot project is the result. Before it is possible to realize more advanced software the bottleneck in existing systems has to be described. It has to do with man machine interaction. Man doesn't know who to comand the machine, and the machine isn't able to obey to the human. It is a bit unusual but Artificial Intleligence can be realized only with better man machine interaction.

IN a teleoperated situation, there is a machine, a human operator and a robot. Realizing these system is possible. The joystick command of the human are converted into machine actions. Such systems are possible for a robot crane but they have failed for more demanding tasks like controlling a train or a care. The reason is, that these systems need very complicated input signals. For example the cockpit of a car provides not a single joystick but the human operator has move the car at the same time left and right and has to decide which speed is needed and in some older cars, the human operator has to control the gear as well.

It is complicated or even impossible to control such a system with remote control. So the question is how to design a human to machine interface which fits to more complex domains. A possible attempt in doing so is a coach based interaction paradigm, which is equal to a hiearchical system. There is coach on the right seat of the car, a driver on the left side, and a cockpit which receives the commands of the driver.

Such a system consists of an advanced command hierarchy. The coach provides the high level goal for example “move ahead”, the driver converts this command into low level steering commands and the cockpit translates the actions of the driver into motor commands. Let us imagine who to automate the pipeline a bit. The idea is to replace only the human driver with a software but let the coach in the car. That means, the human operator plays a certain role. He provides high level goals for the AI Software. The AI software converts these goals into low level steering actions.

Such a system can be realized but it remains an advanced software project. Parsing the command itself is the easiest part of the system. There are many voice recognition software available which are able to convert a spoken command into textual information. The human coach produces a .wav file and this gets converted into the string “move forward”. The more demanding task is the follow up step. How to convert the command string into low level steering commands?

One option would be a lookup table which is realized with a case statement. That means, each statement is converted into predefined action sequence similar to a macro. IN the example, move forward would be converted into “car.speed(20)”. The problem is, that such a direct mapping can't handle different situation. The command “move forward” can be used in many situations but it will generate different low level actions. The more elaborated translation attempt would be to interpret a command as a goal. It is up to the solver to fulfill the goal.

For example, the command “Move forward” can be translated into:

1. remain in the current lane

2. hold a speed of 20 mph.

3. if the current steering angle or the current speed is different, then adjust them

Limits of automation

 

Telerobotis is a well known technique in which human operators are asked to control a robot. A typical example is a crane. Teleoperation is equal to let a human play the game. He has to provide input signals which affects the robot's behavior.

Most robotics engineers are ignoring teleoperation because their goal is to go beyond this capabilities. Robots in the core sense is about replacing teleoperated machines with fully autonomous ones. Let us try first to describe the situation technically.

Suppose there is a video game which is played by a human. The idea is to replace the human with an AI Software. If the software is able to play the game then the human operator was replaced by a machine. But is this assumption correct? Chess engines which are able to play the game are available since the 1980s and they weren't able to replace human chess players. The reason is, that human are playing chess different from a computer. The only thing what is widespread used in the reality are teleoperated chess games which are played over the internet. But on the other side a human is needed to provide the signals.

This sounds surprisingly because from a technical point of view it is possible to replace the humans with a chess engine. But his is not wanted by the chess community. This example questions if it is possible in general to replace human operators with software.

Many programs were written to control robot arms and robot cranes with software. None of them is used in reality. The only piece of software which is relevant is software which supports teleoperation mode, especially algorithm for determine the inverse kinematics of a robot arm. Software which goes beyond this capabilities is available but it is not used in the reality.

The interesting situation is, that more advanced robot projects have the same interaction pattern. Creating a biped robot is possible since some years. But again, these robots are mostly used in teleoepration mode because otherwise it doesn't make much sense.

From a history of automation there are two possible insights available. Simple mechanical machines can be used autonomously but more complex robots are used always in teleoperation mode. The reason has to do with economic goals.

From a technical perspective it is possible to run a robot arm in a fully autonomous mode. The gripper picks up a ball and places it at the goal position. So it makes sense to explain to the audience that this robot doesn't need a human operator. But, if the machine is used for economic purposes, for example in a factory it will be run be a human operator and never in the autonomous mode. In the past it was argued, that the software is not highly enough developed. But the underlying problem has to do with man machine interaction .

What if voice commands don't work?

 

On the first look, voice command seems like a reliable man machine interaction technoogy. The human operator speaks a command like “stop” and the robot is executing the command. This allows a soft transition from teleoperated devices towards fully autonomous one. A possible application would be voice operated train.

The main purpose, why voice commanded robots are interesting is because they are within reach of today's technology. The engineers are able to build such systems and they can be deployed for real applications. In the example, with a voice controlled train, the robo-train would understand 20 simple commands and the human operator remains in the loop all the time. THis allows to equip the system with software without bypassing the existing human operator.

If voice commanded robot are a great idea, it is a here to stay, right? Not exactly. The problem with voice commanded robots is, that they are similar to all sorts of robot an AI Technology. Such technology isn't available yet but it has to be invented. The promise is, that after the train is recognizing voice commands, the human operator can reduce it's workload. But how realistic is such a promise?

The main problem with AI technology no matter which sort of them is, that it has a tendency to work not as robot but as artificial life. It is pretty easy to design systems which are not in control of a human operator, and a voice interface is not an exception. The pessimistic prediction is, that a voice controlled trains won't work. After installing the software and explain to the train driver how the system is working it won't fulfill the needs. And even the engineers who have programmed the device doesn't know the reason why.

But let us go a step backward to understand the situation in detail. The idea is, that a voice controlled train reduces the workload for the human driver. Instead of steering the train with a joystick, the human operator has to talk like “move slowly forward”, “hold speed” and so on. Instead of pushing the joystick manual for 10 seconds forward, he speaks a single command in to the micromphone which produces the same effect. The situation is introduced in the literature as a coach system. The human operator plays the role of a coach. He provides only the general goal and the system figures out the detail. Very similar to explaining how to control a train to the newbie.

In such a case, an experience train operator will talk the same way. He will say, “move slowly forward” and the newbie operator has to execute the task The problem with a robot who interacts in such a way is, that he provides an additional complexity not available before. That means, the robot won't simplify the situation but it will make the things difficult to grasp.

Perhaps it make sense to investigate who a coach to newbie interaction works in the classical setup. The experienced trains operator gives a command, for example, he can say “start the engine”. It is up to the newbie train operator to fulfill the request. He has to press a button for starting the engine itself and he has to make sure that the warning light goes off. In response the newbie can answer the request with “engine was started”.

The interesting fact is, that such an interaction has nothing to do with the train itself, but it reflects the communication between two humans who are operating on the engine. Both operators are forming a hierarchy. The experience train operator decides, that the engine has to be started at a certain moment. And the subordinate has to figure out how to do so.

The interaction between an experienced operator and a newbie is the best practice method for teaching a complicated subject to a newbie. Even if somebody has no experiences with train control he is able to play the role of a newbie who has to execute the commands of the operator. The reason why this social role works well is because a human is motivated to play the role o a subordinate. In contrast a voice controlled software won't do so. There is no need for the software to execute the commands from the human operator and at the end the human has to follow the instructions of the robot.

Whats the problem with self-driving cars?

 

Some attempt were made in the past by Waymo and Tesla to push autonomous driving forward. In an interview a journalist has asked at which moment in the future these cars will become available for the public. The engineers itself doesn't know. There is a general prediction available that after 2025 autonomous cars will be available but it is unlikely that this will become the reality. In theory it is possible that in 2025 the starting point is delayed to 2035 and so on.

This sounds surprising because from a technical perspective all the problems are solved or can be solved with minor improvements. Today's software connects a car with a cloud based infrastructure, a 3d lidar scan creates a precise map of the environment, the SLAM self localization algorithm knows where the car is, and advanced path planning generates the route for the car. So, where is the bottleneck, why are autonomous cars not available?

The problem is, that autonomous cars are without any doubt an AI technology. AI means, that it wasn't available in the past but everything was invented from scratch. The interesting situation is, that Artificial Intelligence has failed to control more simpler domains, which are autonomous trains and autonomous subways. And the engineers want to automate highly complex cars which are driving next to human controlled vehicles.

Instead of asking at which moment autonomous cars will become available we have to ask why it is so complicated to automate cargo trains and the subway. From a technical point of view, a train has no steering wheel but it can only accelerate and brake. The workload for the human operator is much lower and it is easier to create software which can handle the task by it's own. The reason why these software isn't used in reality is because train automation is to complicated for today's AI as well. Or to be more precisely, every task in the world which is important can't be handled by AI algorithms.

Let us analyze some existing press reports about train automation. There are some success stories available about robotics project. The shared similarity is, that all of these attempt are explaining how the future will look like. Not a single example for a working prototype is available. What all the railway companies are doing is that they are figuring out how to realize such future trains. But no one has found a reliable solution which can deployed in the reality.

To understand the situation we have to explain the difference between AI based automation and non AI automation. Non AI automation means to build highly complex machines but let a human operator control the system. AI based automation is the other way around. These systems were designed without a human in mind. Let us give some example. Non AI technology is a minor improvement of existing machines. For example the train gets a better chair for the pilot. Or the monitor gets a higher resolution. Everything else remains the same and no experiments are needed. This sort of technology has a high chance to become reality within the next 10 years.

In contrast, AI based technology is trying to reinvent something from scratch. Instead of improving something by 0.1% the idea is to revolutionize the domain, reduce the costs to a minimum and show what state of the art technology is about.

Autonomous trains

It was claimed that autonomous cars and trains are not possible within the near future. To understand the reason why we have to investigate some press releases about current projects. Autonomous trains are introduced to the audience as a demand. It is explained, that the requirement in the future will become higher and there is need for robotic trains. This statement is correct. Indeed the demand for such technology is there. In another press statement it was written that in the year 2019 an autonomous train was demonstrated on a 48 mile test track. Some detail information about the technology was given too. In another article, autonomous trains are compared with automatic elevators and it was made clear that some challenges need to be addressed before the technology is available.

Does this look like something which is available today? Or which can be realized within the next 10 years? No, it is a vision for future transport systems and nothing else. It is a technology which can improve the life and reduce the costs but nobody knows how to realize such things.

Perhaps another example would help to understand what the situation with robotics is. In the year 2011 the japanese powerplant in Fukushima was hit by a tsunami. At this time, no robots were available to fix the problem. This was a surprise for the public, because in the TV many prototypes were shown in the year before. From the year 2011 until 2020, robotics engineers have taken the issue more seriously and have developed new robots, which can be used for practical applications. Really? No they didn't. The current status is the same like 10 years ago. No robots at all are available. Even remote controlled robots are not reliable.

Why are robots not available?

How can it be, that on the one hand advanced biped robots are developed by engineers and autonomous cars are demonstrated on a real road, but at the same time, the technology is not available for deployment? The paradox can be understood by explaining what is shown exactly in a video or in a paper about a biped robot. Suppose a new robotics video was released in which a robot is able to climb stairs and play football. For the audience, this demonstration looks very impressive and the engineers are proud of the newly written software. The shortfall is, that the demonstrated technology isn't a robot but it is artificial life. Artificial Life looks very impressive to humans but at the same time it is completely useless.

A useful robot is able to replace a human worker. This is the meaning of the word robot. In contrast, Artificial Life doesn't replaces something but it is adding something not available before. Creating artificial life is from a technical perspective an easy to solve problem. The early braitenberg vehicles, a computer game with AI characters and today's biped robots can be understood as artificial life. These machines are doing something, and they need energy for the task. The problem is that humans can't utilized this behavior for their own purpose.

What humans are doing instead is to invent the task which fits to the robot. Let me explain the situation in case of an autonomous car. The starting point is to build the car including the sensor and the software. An engineering team assembles the onboard computer, the lidar and the gps navigation system into a self-driving car. In the next step the car is released to a track on the normal road. The assumption is, if the cars drives on a normal road during a rush hour and no human intervention is needed the engineering team was successful.

No they don't. What the engineers have done in reality is to redefine the public road into an experimental track. Not the AI controlled works fine, but the other cars are robust against interruption.

Kitchen robot

Robots are not helping the humans but they force them into a certain social role. In case of a kitchen robot this problem should be explained briefly. Suppose, a kitchen robot was built which can make a sandwhich with two dexterous arms. Building such robot is possible with today's hardware and software and some online videos are showing how such a robot will look like.

Focussing only on the robot itself is only the first step. The more interesting problem is, how to deploy this technology into the reality. There are two cases available. First idea is, that a normal kitchen gets improved by the robotic helper. And the second idea is, that the robot gets deployed into its own kitchen and humans are invited to join the room.

The second case is shown in the reality. The robocup@home challenge has demonstrated these robots. In all the videos, a newly created kitchen was shown and the human participants are acting with a certain rule set. They get a clear instruction how to interact with the sandwich making robot. At first, the human has to smile. Then he has to press the start button, then he waits some minutes until the robot is done, and then the human has to take the sandwich and eat it. He has to stand in a certain angle to the camera so that the robot and the human is shown at the same time.

The interesting situation is, that the robot can't be deplayed into a real kitchen. Because in this case, the human operator has to guideline who to interact with the robot. It doesn't make much sense to press the start button if no camera is filming the scene. Even the engineers who have created the kitchen robot have no working robot in their own kitchen. If they want to eat a sandwich they will order it from a company.

Artificial demonstrations

Perhaps it makes sense to analyze in detail how robots are deployed into artificial use cases. The robocup@home challenge was mentioned already. It is one of the most famous robotics shows in the world. In a larger building many robots are shown at the same time. There is a soccer playground, a kitchen and a maze with obstacles. All these domains are created from scratch. That means, the kitchen isn't used as a real kitchen, but it is a kitchen to deploy robots.

In such a non human kitchen, the robot laws are established. The engineers who have programmed the robots want to show a certain behavior. They are creating around the robot a scenario which contains of different tasks. What will happen, if no audience and no camera are available during a demonstration? The robot kitchen is useless. It works only if somebody will see what is going on.

This is perhaps the most major difference to a real kitchen. In most kitchens, no cameras are deployed which are filming the scene, but a kitchen is the room in which the meal is prepared. The interesting situation is, that the gap between an artificial kitchen and a real one can't be bridged.

All the advanced robots shown in television and described in academic papers are working fine in artificial situations. Around a certain robot, a scenario was created which has to be solved by the robot. A biped robot needs as an environment a staircase. An autonomous car needs a road, and a pick&place robot needs an object. The funny thing is that the robots can handle these domains very well. As the result a human observers things, that the engineers have solved a certain problem with advanced algorithms. The problem is that an artificial playground a real world applications are working with different rules.

Basically spoken, if a sandwich making robot is creating a sandwich in it's own kitchen the task is useless. Except the engineers, nobody cares about the result and nobody would buy such a robot for it's own purpose.

What is a robot kitchen?

A robot kitchen is an artificial playground to demonstrate AI technology. It contains of a refrigerator, a table and some food and helps to evaluate the capabilities of a robot. A certain robot can grasp the apple from the table or he fails with the task. A robot kitchen is some sort of a game. It is not part of normal life but was created from scratch to evaluate the consequences.

Robot kitchens are used frequently by robotics engineers. All the papers in which advanced household robots are mentioned are shown in robot kitchens. The assumption is, that a robot kitchen is a great testbed which helps to bridge the gap between reality and imagined future. If a task was demonstrated in the robot kitchen, the robot is ready for deployment – this is at least the claim.

What exactly is the meaning of deployment? The idea is that a robot is sold to the public, replaces human workers and is doing a task by it's own. The reason why robot kitchens were built is because to support this goal. The paradox situation is, that at the same time an artificial environment prevents that a certain robots gets released to the public. To understand this situation we have to take a look into the history.

In the 1970s the assumption was that industrial robots can be deployed directly to the industry. The engineers have identified a workplace, and created a robot for this job. Then the machine was powered with electricity in the hope to increase the automation level. After some attempts it was recognized fast, that the newly installed robot was doing a poor job. After a while the machine was switched off and it was labeled as a failed automation project.

The idea of the engineers to improve the situation was to invent a playground to extend the testing of a new robot in a realistic scenario. Such a playground can be a virtual simulation on the computer screen, or a robot competition like micromouse. Since the 1990s, such playgrounds are used frequently and they evolved over the years.

Creating new playgrounds and programming robots which are working well under artificial condition is an easy to solve task. What remains an open problem is how to transfer a robot from the playground into the reality.

Automation market worldwide

According to the latest stats, around 2.7 million robots are available in the world and each year 350000 new robots are shipped to the companies. On the first look, these numbers are looking very positive. It is a sign, that the robot revolution has started and it is growing.

A closer look into the details will show, that automation is never realized with robots but with automation technology. Automation technology means to install electrical servo motors, assembly lines and even an electric driven truck. There are many obvious problems with the mentioned 2.7 million robots worldwide.

First issue is, that a large amount of robots are nothing else but classical automation technology. For example, a CNC machine is not a robot. Second problem is, that lots of real robots for example robot arms are not used in production environment but they are sold as teaching robots. They are used outside the factory in a university lab.

A positive future

There is a reason available why robotics projects are not analyzed with objective measurements. Because the engineers are believing in a vision. They want to change something in the world, and see a robotics project as one way in doing so. The first place in which robots were mentioned was the science fiction literature. They were developed as tool to tell a new sort of stories. Then the same idea was established in the reality. A robot project gives an idea about a possible future world.

If robotics engineers would be honest to themself, they know in advance that their robots are useless. But if they are not creating such projects, who else is in charge? Somebody has to stand up and claim that robots are useful in the world of tomorrow.

Suppose it is possible to convince today's robotics engineers that their work is useless. All the autonomous cars, robo trains and kitchen robot projects would be stopped because of missing values. From a rational perspective this would be the optimal decision but is such a future a nice place to life?

Warum größere Projekte in Python keinen Sinn machen

2017-10-04 C++, import from other blog

Von den Sprachstandards her ist Python ausgezeichnet um größere Projekte darin zu realisieren. Die Python Virtual Machine ist hinreichend robust, die Python-internen Möglichkeiten zur objektorientierten Programmierung sind vorbildlich und das Modulkonzept erlaubt es Klassen zu Packages zu aggregieren. Technisch gesehen kann man mit Python durchaus Projekte mit 100k LoC oder sogar noch mehr realisieren. Es gibt nur ein Problem: wer möchte diese Programme verwenden? Endanwender machen üblicherweise einen großen Bogen um GUI Applikationen die in Python erstellt wurden, und Systemprogrammierer werden garantiert keine Libraries einbinden, die in Python geschrieben wurden. Dagegen spricht schon die geringe Performance. Python erinnert an das Schicksal was Turbo Pascal ereilt hat: es ist eine Lehrsprache in der Programmierausbildung, kann aber nicht für reale Projekte eingesetzt werden.

Die Sprache als solche ist vorbildlich: Python ist sehr elegant designt. Und es lässt sich darin auch produktiver Sourcecode schreiben, in dem Sinne dass man für einfache Aufgaben wie das Sortieren eines Arrays eben nicht erst tagelang in Foren um Rat fragen muss, sondern einfach den pythonic way of life verwendet. Nur, stellen wir uns mal vor wie das in der Realität konkret aussieht. Man schreibt sein elegantes Python Programm runter, es besteht aus 12000 Lines of Code, nutzt dafür selbstverständlich mehrere Klassen und dann? Rein theoretisch ist das Script jetzt überall ausführbar, aber wer will das auf seiner Maschine tatsächlich verwenden? Das Problem mit Python ist, dass es nur eine weitere Programmiersprache ist in einer ganz speziellen Nische (anfängerfreundlich und interpretiert) und das der damit erstellte Code garantiert nicht in größere Projekte wird einfließen. Genau genommen kann man Python Programmierer nur bemitleiden, weil ihre schönen Programme sonst keiner haben will. Java Programmierer werden ganz sicher keine Python Bibliothek in ihr Projekt einbinden, C Programmierer auch nicht. Mit etwas Glück kann man die Library im Pypi Repository unterbringen, aber das wars dann auch schon. Es ist keineswegs Zufall dass es keine großen namenhafte Python Projekte gibt, mit mehr als 10k LoC. Wie gesagt, rein technisch geht das ausgezeichnet, nur leider ist die Welt außerhalb von Python sehr viel kritischer in solchen Dingen.

Ich bin mir nicht sicher, ob Guido van Rossum der Welt einen Gefallen getan hat, als er die Sprache erfunden hat. Auf den ersten Blick hat Python viele Vorteile. So richtet es sich nicht explizit an Informatiker sondern an Wissenschaftler aus den Bereichen Physik, Linguistik und Geschichtswissenschaften. Ferner ist als interpretierte Sprache mit kurzen Edit-Compile-Run Zyklen konzipiert wodurch man in kurzer Zeit viel Code schreiben kann. Genau genommen ist Python also in eine Lücke vorgestoßen wofür es davor noch keine Sprache gab. Aber kann es wirklich das Ziel sein, zu den gefühlten 500 Programmiersprachen immer weitere hinzuzufügen um darüber die Spaltung der Entwickler voranzutreiben? Reicht es noch nicht, wenn Java und C# Programmierer gegeneinander arbeiten? Braucht man neben PHP, go und Perl noch weitere Sprachen? Python hat sogar das seltene Kunststück fertiggebracht zu sich selber inkompitbel zu sein. Bekanntlich laufen Python3 Programme nicht mehr auf einem Python2 Interpreter. Und das Pypy Projekt ist zwar ein JIT Compiler kann aber nicht alle Bibliotheken aus cpython verarbeiten. Irgendwie ist Python eine ganz eigene Welt die im universitären Umfeld prächtig gedeiht und die dazu führt, dass Leute ihre Zeit verschwenden. Anders kann man es nicht ausdrücken, wenn man Ressourcen in den Aufbau von Python Sourcecode investiert.

BEISPIEL
An einem kleinen Beispiel möchte ich das Thema vertiefen. Früher habe ich schön mit pygame Spiele programmiert. Das geht wunderbar einfach, und mit erstaunlich wenig Sourcecode. Man fängt einfach oben an mit “import pygame”, aktiviert das Fenster, und schon kann man seine erste Box auf den Bildschirm zaubern. Jetzt wo ich nicht pygame nutze, sondern in C++ mit SFML das Spiel realisiere ist es deutlich aufwendiger. Man muss sich durch Manuals auf English wühlen, es gibt für alles mindestens 4 Möglichkeiten und mehr Sourcecode benötigt man auch. Für den Computer macht es keinen Unterschied. In beiden Fällen sieht man eine GUI in der etwas angezeigt wird, und beidesmal mit ruckelfreien 60fps. Der Unterschied liegt in der Community die hinter der Sprache steht. Projekt-1 wendet sich an die Python Community, also an Nicht-Informatiker, während Projekt-2 sich an C++ Programmierer richtet. Die Community unterscheiden sich im Anspruch an sich selbst. C++ Programmieren tönen lautstark dass sie die besten Programmierer der Welt seien und demzufolge haben sie auch den Ehrgeiz die besten Programme des Universums zu schreiben, während es in der Python Community sehr viel entspannter zugeht, in dem Sinne dass man sich gegenseitig versichert Anfänger zu sein und überhaupt sich eher mit mit inhaltlichen Dingen und weniger mit Programmieren beschäftigt. Damals in Python war meine Produktivität immerhin bei stolzen 10 Zeilen Code am Tag, jetzt mit C++ in SFML ist sie abgesunken auf 5 Zeilen täglich. Dadurch verdoppelt sich natürlich die Zeitdauer bis das Projekt fertig ist.