Neo-AI

How to Teach a Robot

Finally PSI our artificial, simulated being is born. It does what it's supposed to do. You will be able to see for yourself in just a few days, right here in the Internet.

In the meantime we have more work to do. We have to teach PSI our natural human language. After all, it's just a newborn, able and willing to learn, but still a blank sheet.

How to teach PSI now? The human baby has a lot of time, about two years to learn to speak. We want PSI to learn much faster though.

To start, we are developing a scenario sufficiently sophisticated but not too difficult: We teach PSI baking and cooking virtually. We show it the ingredients like flour, sugar, milk, etc., and tell it what to do with them, step by step. At the same time, we tell (in a human language) PSI, what to do. 

flour	sugar	milk
whip cream	knead dough	spread dough

That way PSI not only learns to virtually bake a cake, but a the same time learns the names of the ingredients as well the names of the actions and how they relate.

After that we will vary the words describing the same ingredients and actions. In the end, we will be able to present PSI the instructions and it can bake (virtually) the cake all by itself.

Going from there we start to introduce more variety, other ingredients and change the recipe.

We'll keep you informed on the results.

Status update

Finally it looks like we are getting ready with the initial version of our AI system. Currently we are modifiying it to a web application. Yes indeed, you will be able to play with it really soon.

What can you do with it? You wil find a very simple game engine with a grid of 5 by 5 squares. Each square will display an image. Clicking on an image will change something in the game. The "game" is actually the virtual reality the intelligent agent "PSI" will "see". You can "show" (=teach) something to PSI by clicking on some of the pictures. PSI learns that. At the same time you can "tell" something to PSI by entering text in a box. PSI sees that as well. Both the pictures as well as the text form the "situation". PSI perceives the situation.

In addition, PSI has motives. It has three motives to start with: thirst, hunger and the need for affiliation. Some of the "pictures" satisfy hunger and thirst. That is, PSI learns, that a click on the picture with the slice of bread satisfies its hunger. A click on the picture with the glas of water satisfies its thirst. By entering an appropriate text like "drink water" or "eat bread", PSI learns that as well.

The third motivation, affiliation is slightly different. PSI likes you and wants to please you. That is, it wants to do what you tell it to. Whenever you click the "well done" button, PSI's need for affiliation is satisfied. That way you can teach PSI arbitrary things in this simple world.

Please be patient, very few weeks only and you can try PSI for yourself.

Year End Update

Unfortunately the development is taking longer than we have hoped for. Yet, there is still hope.
What we have accomplished so far is quite a lot:
We have a motivation controlled system capable of recognizing and memorizing moving images (i. e. a kind of poor man's video) and text. This is the foundation for all coming features. The main problem currently is speed of execution. We had to limit the size of the image sensor to a mere 25 pixels (ordered in a 5 x 5 matrix of true color elements) to achieve reasonable performance.
The plan for the next weeks is to increase speed by a factor of 100 to 1000 or more. This can be accomplished by using a caching approach, optimization using low level languages like C++ or C and multiple threads on multiple processor cores. Even FPGAs come to mind.
Watch out for interesting news in the next few months.

Development Update

After months of silence, an update on the status of our system is definitely due.
Up until the beginning of August we cleaned up and finished the basic architecture of our AI system based on Dörner's PSI-Theory. Then came vacation. Now, back to work.
Where do we go from here? We decided to start with text generation and image recognition.
That is, we build an operator to output arbitrary text. Since it can recognize text already, this is the second half of the feedback loop necessary for learning to understand the meaning of text.
The other, much more interesting part, is building an image sensor. Dörner had an image sensor in his original program, however the actual implementation had somewhat limited capabilities. If all goes as we are hoping for, we will have an extremely powerful and versatile image recognition system.
After these two components, everything is in place to tell the system about the world, teach it "Weltwissen".

Thoughts on Computer Vision

For natural language recognition by computers to work, working computer vision would be very beneficial. Unfortunately the state of the art in computer vision is not at all helpful for this purpose. Computers still have a very hard time recognizing real world objects reliably and under real world conditions.
Prof. Dörner's Psi-Theory contains some ideas for a solution. Dörner incorporated image recognition as part of the perception process 'HyPercept' in the simulation program. HyPercept recognizes horizontal, vertical and diagonal lines of black pixels in simple bitmap images. HyPercept is also able to re-recognize items in images it has seen before as long as they have exactly the same size and orientation. The image is decomposed to lines and their respective endpoints are compared.
How can this approach be improved and made more general? How does the human eye work for comparison?
It is not very difficult to find lines of contrast in an image using current image filtering techniques. This would remove the restriction needing black pixels for image recognition.
Another step would be to not compare the lines' endpoints, but to compare the relative angles of the lines only. Then size or orientation of the objects don't matter anymore. The human eye has sensors to detect angular lines too.
The big problem remaining is the recognition of three-dimensional objects. Technically it is not difficult to project 3D objects on a plane, but is that what our brain does? How could it figure out the three angles of orientation of the object to properly project and then compare it with what is known?
I propose the following solution: a human baby learns both visual object recognition and object recognition by touching using its hands and mouth. The little human learns both visual as well as haptic features at the same time. He is therefore able to correlate form, corners and edges as well as their respective two-dimensional represention.
Is that enough of an explanation? I don't know yet. We will have to find this out by trying it.

Human Resource Management, Maritime Simulation

Last week was an especially exciting one with lots and lots  of things new to learn.
I had the pleasure of going to Scandinavia with Prof. Dr. Stefan Strohschneider to learn about a new training project.

The idea is, to put a couple of people (6 to 10) around a table and have them handle a simulated crises. The simulation, in this case, was a cruise ship sailing from the North Atlantic to Miami. The simulation happens on a computer. The computer prints out messages on paper every 30 seconds on what is going on aboard the ship.  The crew, that is the 6 to 10 people being trained, have received instructions, a map and blueprints of the ship. The crew has decided on who will take what role in the game (captain, first engineer, first steward, ship doctor, etc.). The goal of the simulation is to bring the ship home with as little damage and casualties as possible. The crew gives orders in writing to the instructor, who in turn enters these into the computer to influence the simulation.

The game includes a short time for the players to get acquainted with the scenario in general and then, after they have set up their individual roles, with the task each one of them has to perform in the simulation.

For the game to have an effect, a debriefing session follows the simulation. That is where people are told what they did wrong, could have done better or what was good. Usually the simulation is followed by another simulation the next day.

The people to be trained can be from all kinds of institutions dealing with safety: police, fire fighters, factories, hospitals, airlines or, like in this case ship officers.

Autonomous Robot Going Shopping

Imagine an autonomous robot, four wheels, battery operated, the size of a small baby buggy. Its controlled by a computer, navigation by GPS, communication by GSM or UMTS. It has cameras, laser or ultrasonic distance sensors and carry a payload of maybe 20 kg. The hard- and software, artificially intelligent, capable of machine learning, are at least as capable as the winners of last year's Urban Challenge.
Imagine you head such a machine, you could send it to go shopping for you. It could pick up the groceries you ordered via internet or deliver documents in large cities. It is sufficiently small to cause not too much of damage in case of an accident with a car or a pedestrian.
Technically, this doesn't seem too great a challenge. OK, the cars taking part in the Urban Challenge had trunks full of electronics, but the same computing power will be available in much smaller dimensions only months from now.
It raises some interesting questions too: what if there is actually an accident on a public road with a person or with a car? What if an accident with another robot occured? What if the robot is kidnapped? What if the goods from the robot's cargo area are stolen? May the robot defend itself with spraying water, pepper spray, with high voltage or with a Taser?

New Free E-Book on Dörner's Psi-Theory

I'm proud to announce the initial version of my free e-book covering the essentials of Prof. Dörner's Psi-Theory. It is an alpha version yet, but still quite informative for those who cannot read Dörner's german books.

Conscious Robot Is able to Improve Behavior on past Experiences

This was a very interesting week. On friday we visited Prof. Dörner in his "lab". He is currently using a version of the Psi-Simulation software with several dozen of beings (mice) in an island environment. He studies the beings' social behavior under, among other things, the influence of alcohol (simulated).
He now added an, at first glance seemingly very simple, as it turns out however, extremely powerful feature. Its a graphical tool to analyse the mice' behavior protocol. The protocol analyzer shows visually the actions taken by a mouse and its feelings. When looking over the protocol it becomes very easy to spot areas of unexpected actions and behavior. The tool can be used to find possible program errors as well as study the behavior of a mouse in a certain situation. Besides the actions taken, curves showing the values of emotional parameters are drawn in timely correspondance.
It is therefore very easy to find causes of emotional extremes by just looking at the protocol. One can simply see the reason for the mouse being sad or angry by looking at what as happened to it a few cycles before.
It is not very difficult to imagine an add on to the current mouse simulation, that is able to analyze the protocol by itself and have the mouse improve its behavior by itself - the robot (mouse) learns from experience and improves its behavior.

Advanced Magic to Develop Bridge Simulator for Training Officers

It looks like we are going to develop a simulation system for training ship officers. With know how and in cooperation with the Department of Intercultural Communication of the University of Jena we will write a ship simulator to be used in the training of officers for freighters and tankers. The simulation will be a so call mid-fidelity simulation, where the computer is used to simulate a ship on the ocean. No 3D-effects are planned, rather the simulating computer will spit out plain text messages as to what is occuring onboard.

In the course of the simulation problems arise on the ship and the crew, that is the participants of the training, has to cope with the difficulties. The focus is on human factors in cooperation and behavior under stress, not on the technical difficulties. In marine speak the training is called bridge resource management.

New regulations from the International Maritime Organization may require this kind of training in the near future. The mid-fidelity training has the advantage of being much cheaper than training in true ship simulators. Also the quality of the learning process is improved since no high technology equipment diverts participants from training their social skills.

In another scenario, we plan to develop a simulation of chemical plant, a hospital, an oil rigg where a chemical spill or a fire occur. The participants need not be familiar with the actual technical problems, since emphasis is on cooperation, leadership and behavior in criticial situations.

A special value for us as developers of artificial intelligence software is, that we may learn more about how people behave, so we can refine our systems for natural language processing.