AI and Gaming

I must agree... AI is certainly to be done in stages.

Determining what an AI is meant to do is difficult, as well as determining how it should be done.

I think that a better method would be to use some sort of automatic testing system, that would allow an AI to fix itself. This is where neural networks are helpful, as there are algorithms that automaticaly adjust the neurode weights without human interferance.

Here, with AIP, Chuck is working on creating an AI that can be taught, not only to do one specific task, but several related, or unrelated tasks.

Z.

Hi,
Z, when the program is ready to post I will do it in the Communications area as you suggest.

CodeRonin, do not think that all of us are like Zaei and are brillant programmers. ;-) My study of AI has been theoretical in books that I have read. Lots of ideas but not immediate application to programming. I did get a book from the library called "Practical C" by O'Reilly & Associates. These decision trees are explained in very simplistic terms so much that the lights came on today :-)
It was a simple figure but it all became clear. Now I know what it is. Can I apply it? Not sure.

I am going to perform some nonsensical tests on my AIP. The most important will be lying to my program as the teacher. You see, right now, the program is 100% trusting. I am curious how much I can mess up the 'little fellow' by lying. Of course 'lie' detection would come in later program versions.

Back to debugging my file writing and opening.

Regards,
ChuckB

Of course it's hard to determine what an AI should do - but every design is hard.

However, AI's are in general better if they are designed for a specific task.

Although my philosophy is to hardcode as little as possible, when it comes to AI, speed is essential, and hardcode is way faster than any external data exchange, no matter of what kind.

Of course, you can write small AI's by teaching them everything, but as soon as it come to bigger projects, much more hardcode is needed, especially in such slow languages as VB.

AI?
Hardcoding AI for specific tasks is not really my area of interest. I have written many games with 'AI' in QBASIC years ago. They worked fine. My chess game played legal chess, though my young children could certainly beat it. The problem with my chess program was simply this....

...it could only play chess!

So, my concept of AI goes beyond the performance of a single task but rather sees an engine that is capable of handling several AI tasks...with the possibility of persons adding modules to extend its engine capability such as 3D coordinate systems, time, comparisons of objects (size, color), natural language processing, etc. The modules could be added using just about any code as long as the various libraries or modules obeyed certain conventions when passing arguments, etc.

Imagine, you want to add an agent to a game. You define its characteristics, its intelligence level, mobility, etc...into a script file. The script file along with a module is linked into program and voila...you have added a worthwhile contender to your game just be selecting a few attributes. Of course the game program must have the code to handle time (time compression), coordinate systems, interface agents with virtual world, etc.

So, this module (DLL or LIB for example) has the ability to play a chess game or navigate around in a virtual world and seek to destroy the human player and so on. It accumulates knowledge and saves it into a database for later retrieval. It then locates related AI agents like itself and shares knowledge.

Far fetched? Maybe. But then it is just a dream of mine. How to implement it? Not sure.

So, my AI program will handle simple fill in a grid type games. Then hopefully expands to play checkers, connect 4, chess. Then hardcode capability to play Risk, Monopoly, dominos, card games, etc. The list goes on.

Teaching time should be very quick...like teaching an adult to play chess....they learn quickly...and with some practice get quite good. There is something about the human brain that allows it to learn a variety of games. All of these facilties come to play in real life with humans...therefore AI that can play games I believe has a chance of being used outside of the gameplaying arena.

So much for my thoughts. ;-)

Time to code.

Regards,
ChuckB

Chuck, what you've described would be the optimal AI : short learning time, little hardcode and availiable for all tasks - true AI.

The only problem is: No one can be sure if it's even possible to create true AI.

Of course, it's great to do researtch on true AI, I'm glad people are interessted in creating true AI - I am, too, and I'm doing my own studies in this direction.

However, if an AI is NEEDED for a specific task, it's nonsense in my opinion to teach it hundrets of other things, which it'll never need - and lose time in this way.

As you know, you don't have time to do everything perfect in a big project (when it's not just for you), so you'll need to find a middle-way.

No company will allow you to spend loads of money and lots of time to develop an AI to play chess - for a good reason.

So, for own study there's nothing wrong with searching for true AI (as I do, it's my holy Grail ;) ), but when it comes to an AI that is needed to play a role in a project, you won't have the luxury of implementing true Ai if you want to finish your work.

It's fine to have someone to discuss with,
CodeRonin

Hi CodeRonin,
I agree with overkill on writing task specific AI. But consider this.

There is a program called Cyc which just celebrated its 18th birthday. Led by Lenat this program takes natural language input and is capable of applying 'common sense' to many things. For example, a dead dog cannot play, falling off a steep cliff could result in severe injury, etc. I understand it is available on line for training by internet users.

Lenat's thoughts are to take this package which fits on 1 CDROM and to put it on future Operating Systems for some purpose...

My concept on AI would be to produce a module that can be taught <once>. The module and the knowledge base would then be used as one unit for future programming needs...an AI library I suppose.

What really interests me is that with all of the predictions about what computers could do in the future, it is the field of AI that has been overly rated...i.e. computers that could beat the best chess players as early as 1964, fully conversational type programs by 1990, etc.

My readings on neural nets and all they promise have not lived up to most expectations. Expert systems require intensive and expensive training sessions. So the AI field is wide open for us to explore.

I spent nearly 40 hours programming during the past 5 days on my 'simple' AI program. It is kicking me in the tail but I smell victory on the horizon. :-)

Regards,
ChuckB

This is off subject, but imagine for a second that someone made a fully intelligent AI, with the ability to be smarter than anything on earth. What happens when it realizes, as it must, that it is the smartest? This is an ethics question, because any true AI won't have any restraints or blocks on it; it would be able to learn to hate it's creators. And obviously, with the way the internet works now, it could easily infect every computer in the world. So do we really want to have an AI that is capable of learning anything, as real intelligence is?

Not really, snakeeyes. Though movies like Terminator seem to put AI in a bad light, think about it this way. If someone were to develop a self aware AI, and it took over the world, can we honsetly say that the person who created the AI is responsible? That would be like saying that any mass muderer's/conquerer's/evil dictator's, etc parents are responsible for their actions. Are we to say that procreating is a bad thing, because it might spawn something evil?

Though, if you really think about it, a true AI would not really crave power, domination, etc. At most, I think it would be self preserving (a la, The Iron Giant), and not a tool of evil =).

Z.

Hi Everyone,
I am sure you appreciate seeing a challenging program work after you have spent over 100 hours designing and writing the source code.

Today at 3:30 EST AIP v0.1g kicked 'butt'! It took just three tic-tac-toe games for AIP to learn how to tie and beat my son. My wife created a 5x5 grid game in which you had to connect 5 blocks (cells) in some connecting pattern. Within five games AIP (pronounced 'ape') was playing agressively with what knowledge it had learned and with an unquenchable thirst for even more knowledge. Okay...I may be exaggerating just a bit...

Man! The engine requires data to be formatted a particular way. The actual AI engine is fairly simple...blew me away. I expected tons of code to make this program 'hum'!

I have 5 cosmetic bugs to correct and I will be ready to release v0.2a to the public. :-)

Sorry about the excitement. It's my first real, genuine AI program that wasn't hardcoded to play specific games. Hey, even my wife the schoolteacher thinks its really cool!

Got some debugging to do.

Regards,
ChuckB

Good going! Congratulations! I would have expected that it would require more games for it to start tieing and winning, though =). Good job!

Z.

Hi,
I have started a new thread in the Communication Area supporting my AI program, AIP v0.2a. The thread is http://www.vbforums.com/showthread.p...hreadid=185539

The zipped file - source code, graphics and HTML help pages can be downloaded from http://www.clg-net.com/ai/aip.htm

Regards,
Chuck B

I don't claim to be that up on the theory, but I've been reading this thread and playing with Chuck's AIPs and it seems to me that our man here has created a simplistic neural net.

Each square of the board is a node that gradually recieves a 'win weighting', that allows the AIP to make good decisions about where to play in order to win, or to prevent itself from losing.

That is essentially a neural network is it not?

Chuck: I took a look at AIP.2a, but it wasnt very friendly to my Virtual Computer =). I will take another look later on my main computer (I wasnt sure if you had included an exe or not, which is why I tried it out on my virtual one first (VB doesnt run over a network =( )). The interface looked very nice, but once I got into a game, the computer essentially locked =(. I would put this down to a fault in the VM and/or VB itself, since several of my programs do the same thing =).

Judd: Though the program does assign a weight to each square, it also uses patterns to figure out its next move. AIP appears to try and find the square that a) control has yeilded a win, and b) fits into a pattern that causes a win.

Z.

The ZIP file does include EXE and source code. However, it requires two controls...flex grid and tabstrip/status bar. Maybe this is the problem. Can you run any VB programs..non EXE mode...with your VM

If you cannot then you probably need the two OCX files. I have them zipped at http://www.clg-net.com/ai/aip.htm for those that need them. Of course, they need to be registered with regsvr32.exe I believe...not sure about VM's.

I have received feedback from person online and locally. One report played two instances against each other to see the outcome. The human served as the mediary (sp!).

I am advancing a theory now that will increase the 5x5 limit to an almost indefinite size limit by using arrays instead of 32 bit long integers. I have worked out all details on paper to get v0.4 to play checkers...accept two things.

1) How important is history...is it really needed once AIP needs rules. I think it is a possible disadvantage that AI programmers rely upon entirely too much...modeling humans...for example, don't remember much history in great detail...just a few key facts.
Yet their pattern recognition is superior to programs with tons of historical data....

2) Matching winning patterns is one thing...matching legal moves with board layout...while marching towards a goal of annihalating the enemy requires a lot more thought...and programming

You see humans must be taught rules, but they also must learn strategy...which I must be able to define in my game .AIP file and inside the AI Engine.

Judd,
I wouldn't know a neural net if it hit me in the head...but its a flattering thought that I may have stumbled upon something so cool. :-)

Regards,
ChuckB

The VM is essentially Windows 2000 in a little window =). Everything is exactly the same, otherwise. The application runs, and works simply fine until I actually start a game. At that point, the VM becomes unstable (the VM's fault, not AIP's).

I posted down in the Communication Area as well.

Z.

Hey
well I am sure you want to concentrate on the AI part first and not on the interface. But I think you could improve you AIP a lot by upgrading the interface. What I am thinking of is: Setting up a connection between two of your programs (most likely with winsock) once your programm knows all the rules you could provide your file storing those to both programs than connect them and let them play for a couple of hours if they don't start repeating the same game over and over they are very likely to develop strategies (I haven't seen any of your code, neither am I sure if I would understand it..., so I can't say if they are really going to learn by playing against each other, but I think it is probable)

Hi,

You mention Winsock. I have this question. I can write two programs that can run on the same computer that communicate with each other in two ways. 1) They read/write to a common file, thus they leave and write messages for the other to use.
2) I did some DDE years ago on a very limited basis.

Can I use Winsock to get 2 programs on the same PC to communicate? If so, does anyone have some sample code that I can play with because I would really like to implement this? Can I then use this same code for network/internet communication by specifying IP addresses?

Computer vs. computer?
There are some non-random things that I do in the v0.2a program that results in some eventual predictable play by AIP. I will change this in the near future.

Before two AIP programs play together, I would suggest the human teaches the first AIP program to play a game. After the training session is over...say 10 games...then you allow the two AIP programs to play. This way, one program teaches another to play the game. The emphasis here is on teaching a program to play a game.

That is the current limit on AIP. Once the program knows how to play a game it will do its best to win and to oppose defeat.

M, consider this. With the AIP engine, it is possible to create a new computer game without any programming in the conventional sense..you know, using VB or C++. Instead, we can create games by playing the game and teaching it as we go. This game is saved. We can write another program that has no teaching interface but is a simple interface that allows game play only using the files that were created with the AIP teacher interface.

Now, I think this is how the human brain works...which is why I consider the program AI.

Regards,
ChuckB

Z and others,
Here is a concept question for you.

I know how to save history and rules in a text file for reading into a program so it can operate. AIP v0.2a does that. However, I am not satisifed with this one major defficiency. AIP CANNOT add to its program by writting more programs or program modules...eventually growing to a much larger size. The brain seems to constantly update its 'program' as we learn more and more...so I have not really achieved AI as I see it.

Do you think the following process will work and help me to close up on this.

I create an AIP program module that allows me to read data in from a specific file format. This is not the current game files I am using...but a deeper and more permanent method of allowing self-programming.

'sample file
GAME.ROWS | integer | Method=DefineGridRowSize
GAME.COLS | integer | Method=DefineGridColSize

Now, inside the VB source code, there exists no hardcoded global variables for defining rows and cols in the grid or any other game parameter. So, when this file is loaded by the program it constructs a user defined array to handle these variable types. Something like this:

'define user type
TYPE MyArray
a AS INTEGER 'define all possible variable types here so type is universal
b AS LONG
c as BOOLEAN
.
.
h AS STRING
method AS INTEGER 'stores pointer to a hardcoded method
END TYPE

'update array
REDIM PRESERVE nGlobal(UBOUND (nGlobal) + 1) AS MyArray
IF nVariableTypeReadFromFile = "integer" then
nGlobal(UBOUND nGlobal).a = nIntegerValueReadFromFile
ELSE...continue all variable types after this
END IF

nGlobal(UBOUND nGlobal).method = nMethodTypeReadFromFile

So, now, I can create AIP scripting language keywords...assign them a variable type an associate them with a Method. The problem is that this is a bit pointless since I am still hardcoding Functions or Methods.

Now, am I looking at the ability of AIP to write its own macros? Its own methods? Am I writing my own programming language that can be used by AIP? Am I way over my head? ;-)

Any thoughts would be appreciated. My task seems to be able to write ironclad engine that can process a sophisticated set of keywords and macros...am I writing an interpreter? Macro recorder?

Regards,
ChuckB

well for the winsock question:
yes it is really easy to program a program that can communicate to another one over winsock no matter if it's on a local pc or on the network or the internet.
It is much easier to do it for one pc though since you don't have to bother with the possibility for data coming in slowly..
all you need to do is put a winsock control on your form and set one of you programs to listen (server) and the other one to connect to it (client)...
all of this requires only a few lines of code that I have to look up in the msdn every time I use it...
as far as I know there is a winsock tutorial on vbworld so you could use that code for reference.

I read through most of your thread and I think I understand what you are doing. What I meant is this:
First the human teaches the game to the program. Now you get the game file if I am not mistaken. Copy it to the other intance of your program. Now both have the same 'rule set'. Now you don't have to bother to teach the second program from the first, but you(respectively your program) can go to develop strategies...
if you got enough random actions in your AI it should gain huge knowledge about how to move by playing practically against itself.

...well while writing this I was thinking: You could program your AI engine into a server kind of thing. Now a client can connect tell it where the game file is, send the current state and your AI server will send it's proposed move... this way you would be really flexible, you could have some client that looks much like your current gui that can teach than others that can just play one game that comes with them as you AI file thing..., this way everyone could code a graphical interface for some game he taught your ai first...

Chuck: Essentially, teach AIP to program itself? Actually, I have realized the same thing about our own brains, though not to the extent that might be imagined. I think that our brains only add data processing capabilities. For instance, before we start school, we dont know how to read. At that point, when we learn to read, we learn how to process written information. Once we have that, it is only a matter of expanding a knowledge base (our vocabulary) to learn to read. The same goes for essentially anything. Math, programming, etc. When learning a second language, we simply train our brains to map words and phrases to words and phrases in our original language, in the first stages, and then the second language gets its own knowledge base (fluency). For instance, when learning a second language, it is very difficult to read a word that is unknown and figure out its meaning from context, while it is usually VERY easy to do in the original language.

Basically, what you propose is to write an engine that is capable of not only learning, but also modifying itself to be able to understand more concepts, in the way the human brain works. Conceptually, brilliant. Practically, itll take work, but it seems feasible =).

I would personally be interested in pursuing something like this =).

As to winsock. Everything MisanThrop said is true. That is the reason I am using winsock for communication in my database system, ease, and flexibility (simply point the IP address to some place with an AIP server, and it will work).

Z.

Hi,
I took M's advice and found some tutorials. I created to projects on one PC...added text boxes and connect/send buttons. I got it to work. Even added text to speak control and one line of code so we could send 'spoken' messages to each application. Geeky for sure. ;-)

I kept getting some inpredicatable errors, especially if I 'reconnected' during a session...stuff like that. I couldn't find any solid examples with bulletproof errorhandling for two Apps on same PC. I'll look for that when I have some free time.

I will keep this interface method between 2 Apps in mind as I grind away at the new AIP version.

Regarding AIP 0.4v, I am rewriting most of the code because I have since come up with better ways of dealing with things. I am allowing for 15x15 grid, with larger cells and fonts. I have global variables that allow for 2 types of checkerboard patterns, random patterns where teacher selects cells to be out of play (crossword puzzle looking pattern) and a few other whistles. Now this all works...pretty cool looking.

At www.clg-net.com/ai/aipsl.htm I am composing a scripting language specification for use with v0.4. It is a work in progress but represents the new language...increasing AIP keywords from about 20 to 80...big jump but necessary for AIP to learn chess.

On paper, it is a simple matter to write AIP to play TTT or chess...that is, it plays legally. However, I must spend time coming up with a novel scheme for AIP to take chess lessons and convert them into some principles to guide later play. I have constructed some Excel sheets with checkerboards and 64 bit table to play with '1's and '0's. Perhaps I can get this. Thought getting AIP to learn to play legal chess is my core objective...the rest comes later.

Now, regarding a program that can modify itself...wait for v0.5. ;-) Yeah right!

Actually, the Engine must include hundreds of functions that are hardcoded that deal with numbers, strings, 2D, 3D, etc. From this the program must allow the user to define new keywords, their variable types and functions that they can connect to...or allow for the writing of macros by the program itself...based upon core language set. The Engine would be .EXE/.DLL but the added modification of program would be perhaps ASCII file formatted.

Maybe we can construct a simple program experiment where we define some AI, a goal, some key functions, program self-macro creation ability and see what happens? An example would be to write a program that can add and subtract. Give it ability to write macros...give it comprehension of summing. See if it can learn on its own to create a multiplication macro function on its own...and then use this to discover division, averages, etc.

I think that would be a decent experimental concept. What do you think?

Regards,
ChuckB

Sounds good Chuck. Winsock is pretty easy once you get the basics down (ive done a simple web server, and telnet server before).

Like I said, the self modifying AI sounds like fun, and I think it is doable. I would love to work on something like that with you. The summation and subtraction AI should certainly be feasable, and would be a great proof of concept.

Z.

I've been reading through this thread, and the ideas shown here are genius! Chuck, AIP works marvelously! After about 5 games AIP was winning and tieing every time! This is fabulous! :cool: :cool: :cool: :cool: :cool: :cool: :cool: :cool: :cool: keep it up!

Thanks Steve. I appreciate the feedback. If you can give this a try... Start the program that has been trained by you. Start a second program that has not been trained. You will need to serve as the interface between the two programs. Now, play them against the other.

Theoretically, the second program should learn after a few games. At which time, there comes a learning plateau where the program kind of hits a 'mental block'. :-)

What should happen, but doesn't, is the AI program should do something novel...unexpected...creative...such that it invents a new strategy all by itself...and continues to kick tail.

Of course, this is really limited with TicTacToe. However, I believe this should be the expected behavior of AIP playing a chess game or something in which game strategy can be developed.

Well, if you can do the experiment above, give it a shot and report back to this thread your findings.

Thanks again,
ChuckB

Hi Z,
I am going to build AIP v0.4 to be able to learn legal checkers/chess and larger pattern matching games...with minimal input on further reasoning and creating additional strategies. This will satisfy my requirements for a program that can be taught a significant variety of games.

I will be spending the next few weeks with a pad of paper while sitting with family in evening, attending meetings, sitting at desk between classes,etc...actually working on some model for AIP developing and refining its knowledge base...I am seeing tiny grid cells and bits in my mind as I close my eyes. :-) Maybe I can develop a function(s) that will allow AIP to play better and better rather then just legally. Again, I must avoid any hardcoding of program to play chess...just give it the ability.

I would like to work on the self-modifying AI. It will take a lot of work up front developing a model in concept, specifying goals and the test for success, etc. Of course, this would probably be many weeks/months in the running. I must devote a lot of evenings to developing an instructional video with my PC and that will eat into my time...and AIP v0.4.

So, if you want to start slowly on this project I am up to it. Let me know if the timetable is okay for you.

We can develop Word docs describing our ideas and plans, or use this forum thread or I can setup a web folder and give you FTP access so we can post our ideas, sample code, etc.

Regards,
ChuckB

The idea of making program develope its own strategies can only be done by allowing the strategies to be developed. Giving the program a rather complex state doesn't boost its capabilities, it boosts its uses, often including faulty uses. The artificial intelligence itself can never be improved, its a turingmachine.

But, for those who don't know, it always looks impressive, there's just no magic involved..

I think functional/logical languages like prolog or haskell is rather suited for what you are doing.

Kedaman is sort of right - but I think his expectations are a little high ;) . Chuck's AIPs are never going to 'learn' anything outside their ability to learn things, which will be defined by Chuck. But the same can be said for us - we can't learn to fly by flapping our arms because we are not equipped to do it.

Chuck is the (thinking of) writing an Engine which will act as the universe for AIPs (essentially defining the 'physics' of their world, defining their abilities), just as our universe defines our abilities.

I think what Kedaman is pointing out is that, for example, Chuck will have to 'tell' the AIPs under what situations they will make a random 'unpredictable' move. Chuck isn't going to be able to develop a machine that thinks for itself if he defines its abilities. But I can't see how to write a computer program without telling the computer what to do ;)

Chuck - you were right earlier in the thread about history of games. It's not important if you're seeking to emulate human behaviour. If you considered how you play chess for instance, you are not analysing all moves up to this point, you're generally thinking forward from the current point. (i.e. what am I/is he going to do next, given the current position).

Judd you have understood me correct and made my point a lot clearer i think, thanks :)

For instance, a program that would develope strategies for a tic tac toe game, would be equipped with the statespace of all possible moves (reactions to all possible situations in the game) and to enforce good moves, you define static rules that gives value to each possible move. Now the simplest algoritm would be to iterate trough the best moves while playing, and start valuing moves that defeats the computer in game.

Chuck wants to go further and make the program define its own strategies conceptually, this is not at all impossible, as before you just give it the statespace to define all possible concepts - but this would be rather done with logical languages, there's no better tool for a declarative approach. The problem is though how to set the rules for good strategies, empirically testing could take ages - like evolution.

Which is why I think Chuck seems to have cottoned on to the method of getting two AIPs to play together (takes the tedium out of training) (Chuck, I see no reason why a given EXE shouldn't have more than one AIP personality - that would ease the comms problems!)

I think it would be interesting to teach one AIP checkers (for instance), then have it teach another AIP and see what strategies develop - and see if both improve as players. Perhaps the 'player' with the better teacher will always win??

The hard part is programming the 'impetus' - or the 'instinct' (desire?). To a human, the desire to win a game is sufficient to 'cause' us to develop strategies - I don't think that would be sufficient for a computer.

Even with Chuck's current pattern matching AIPs, which seem to learn your strategies, it comes down to simple rules:

1) Can I win on the next move? Yes, then win.
2) No, can my opponent win on the next move? Yes, then block.
3) If I haven't moved already, make the first 'best' move according to my 'win' matrix.
4) If I can't decide, make a random move.

This does however, work brilliantly at simulating intelligent play (probably because that is more or less the way a human would behave).

Once you are out of the pattern matching arena though, it starts to get much more difficult to simulate human behaviour.

Chuck - when you get your checkers engine running, an interesting test would be to teach an AIP, then have it teach another AIP BUT have one or both randomly make unpredictable moves (i.e. randomly decide if a move is going to be random or 'thought out'). I'm not sure if this will lead to stategies developing, but it could have a noticeable effect on learning times...

Misantrop, I'm not exactly sure what you are asking for, but i'll try to answer anyway with what i know.
There's several interpretations of Artificial Intelligence, the branch in computer science that is exploring how computers can act with intelligence is the broadest, but also a bit confusing. Many thinks its about emulation of human intelligence, which might be just a goal in itself, but not nessesarily what makes a machine artificially intelligent. The confusion is the more philosophical issue what to consider intelligent. I have decided to skip philosophical involvement by assuming that intelligence is the vital information processor for finding a solution to problems, thus making it subjectively a tool for the situation, which is of course most practical - Something is intelligent if it can solve my problem. A turing machine can solve all computable problems, in practice all possible solutions to all possible problems, as long as information is what we seek, and once again to avoid philosophy, there isn't anything but information that could make any difference. Whether, you equip the computer with legs, arms and eyes, or integrate your neural network with its circuits isn't changing the information flow between you/your computer.

Some may thus say its intelligent in a context, to me all software is artificially intelligent, I woudln't call anything software in the first place otherways. They fill a function and thus by processing information they solve a particular problem/subproblem

Chuck, Illusion is a tricky word, if there was something that you couldn't reveal later to be something else, it wouldn't be an illusion either, was a mistake of me to use it. Self learning is a word assuming that self is both an information container and processor, in computer science we see this as Object Orienter Programming, a paradigm which puts information in only partly transculent boxes. I'm one of those who think this is a very faulty and dangerous way of programming, the big unknown causing bugs in all programs designed using OOP is the private state that is contained within applications making them act mysterious and spontaneous. Thats an illusion for sure because they do act according to their program, but partly out of control for users as they are granted ownage of their private state.

In declarative languages, the application has a given answer for every question, for a single input you have a single output. The program does not contain any state at all, so the user owns all the information including what requires the program to act as a statemachine. This is how intelligence should be structured, as it can never be unpredictably hostile as in OOP and casually hazardous due to bugs hidden from the programmer.

Self learning to the preserving external state is like centrifugal force to centripetal (in physics we say that centrifugal is an illusionary force) if you compare to human beings, you could say that your memory is part of your senses.

The other thing, although I would prefer beating up other players in UT or even bots (they do act intelligently) to release emotional preassure, well depends of course on the situation, If I talked to my girlfriend instead I would mostly get a lot more satisfaction. That kind of purposeful illusion makes sense, you wouldn't want to play solitaire on your comp if it showed you the cards. In the end we want humans to be happy, not provide them with eternal truths that would probably make them very sad.

Having the "learning" bots playing against each other sounds good, in fact if you experimentally had two human players play against their own bots to train them, and then have them try to beat each other, for a good while, I'm sure the humans would find themselves more oftenly beaten if not even inferior to their bots when they would try to continue their training.

You mentioned desperation, which made me think of something funny, if you had a program that would continously fail, it would gradually loose its resistance to randomly change state in order to evolve, and as a contrast if the desperate bot actually succeeded as a result, it would "feel very good" and add extra bonus points to that strategy. Whats wrong with having egoistic bots, if they're working for you? :)

Chuck: I think we could get a machine to learn to multiply. There are really only two problems. The first is creating a scripting language that is expressive enough to allow the engine to acheive a goal. The second, and more pressing is actually getting the engine to understand the scripting. Essentially, we need a way to allow the engine to store concepts. Even without that, I could probably write program that, given a certain number of operations, could generate a macro to do multiplication, using a search tree method. The problem there is that as the engine is given more operations, the search space becomes immense, and thus, unsearchable, except for the simplest of problems. This leads to a third problem, that of finding a method of solving a problem, given a knowledge base, and a goal state. Not simple problems, but they can be overcome =).

Z.

Z,
I have given thought to the problems you have identified. Absolutely some problems that seem workable to me.

To gain some insight into the matter I conducted an experiment with my wife as the guinea pig. We used some math manipulatives that my wife uses to teach children to MPY, DIV, SQR, etc. I took 8 blocks and put them on a 10x10 grid.

The rules of the experiment were expressed to her as follows:
1. You can count 1 to 100.
2. You can add to produce sums.
3. You can determine row/column position of all blocks.
4. You can take x number of blocks and move them to fill any pattern....any arranging.
5. You can express symetry (i.e. 2 columns with 3 blocks in each column is symetrical).
6. You can express equality (i.e. 1 + 3 = 2 + 2)
You get the idea...

She placed 8 blocks on the grid. I asked her to move them around into patterns using above general rules. She came up with quite a few patterns. Here is a sample:

4 rows with 2 blocks each. Row=4:Col=2. Defined in a AI generated language(script) I wrote it down on paper as:

1+7 = 2 + 6 = 3 + 5 = ... = 8 = Row(4) OP1 Col(2)

AI can develop the last expression above which we would translate as 8 = 4 x 2. Now AI wouldn't know of the label that we associate with it so it uses OP1 as an operation.

Now this concept must be scripted into a text file as you indicate so that will be a definite challenge.

More operations may be something like this:

Pattern(12) OP2 = Pattern(38)

Meaning grid pattern representated by bits is equal to 12. When flipped 90 degrees CW (OP2) it produces pattern defined by 38.

Pattern(33) OP2 Col(2) = Pattern(9)

Meaning grid pattern when shift to the right two columns presents bit pattern equal to 9.

Is this useful? Not sure. But it would be derived information by AI and has value in that fact alone.


Proposed Interface
=============
Now imagine this program has a calculator GUI. It has three text boxes and two buttons (part of button control array).
The calculator is fully functional for three operations: +, - and =. Now, as AI defines the operator above (x), it adds another button and places the caption "OP1" on the button. As soon as this happens, the user can use that function on the button. Now the new knowledge is saved into a file and retrieved each time
the AI calculator problem starts. As teachers, we have the ability to assign new labels to any new math concept created by program. So, the program can now display X instead of OP1 for our benefit. Inside the file, it maintains OP1.
Two grids could be added to show the patterns being manipulated and their outcome.

The brain sees lots of patterns with 8 blocks. However, only a few actually have useable mathematical applications. Others are just interesting patterns. How would AI discrimate between any useful and non-useful stuff? Don't know. Haven't thought that through.

AI Engine must possess sufficent hardcoded functions to define pre-programmed rules. The macro language must be able to define new operations in terms of this pre-programmed language set plus any 'externally' stored keywords or macros.

Now, multiplication will be just one of potentially hundreds of interesting discoveries made by the program. It may not be the first. However, AI programs normally run out of new rules or operations to create after a short while.

To extend its life and usefulness...it could allow for teacher input to teach new concepts based upon what it has learned...allowing it to go looping around and making perhaps more discoveries.

Since I have TTT on the brain, I see this problem as a problem involved with working 2D grid systems. I think AI must be armed to make and record observations about the grid that is storing this bits representing blocks.

Try this. Grab a checkerboard and place a black and red piece somewhere on the board on different cells. Write a list of observations. Start with general observations that a child would make. Then work towards more specific details. For example:
Progression of detail
1: There is a red piece and a black piece on different cells
2: The red piece is above the black piece.
3: The red piece is 3 rows above black piece.
4: The red piece is diagonal to the black piece.
5: Etc.

I think AI must possess basic ability to denote and record these simplistic observations if it going to eventually MPY.

Regarding the search tree method, you are far more familar then I with that method. You know I use bit patterns and comparisons with AIP simply because I was comfortable with that and because I think that method closely approximates human brain at work.


Regards,
ChuckB

Hi Everyone,

Here is a "bit pattern" question.

With grids up to 5x5 I used a LONG integer to store bits using the lower 25 bits out of the available 31 bits (excluding the nasty sign bit) to represent each cell.

Now, to express grids of 15x15, I do not have a SUPER LONG integer. Three ways come to mind however. Please suggest the best and most practical. Are there any other schemes to consider?

1. Grid (15,15) Two indices (row and col) define location.
2. Grid (1 to 225) One index defines cell...using position 1 as the top-left cell.
3. Pseudo SUPER LONG integer using a string to hold at least 225 (15x15) characters.

I have used method 1 above in other game like programs. I must admit, defining patterns meant a whole lot of IF this and this and this and this and this...THEN something. Hardcoding was really tedious. That is why using LONG integers was so nice and easy.

Method 2 more closely matches what I did with LONG integers. I can cloak the arrays into functions so they seem to operate like very long integers. I would have to reduce logic operations such as XOR, AND for example into individual bits...one at a time. Better though then method 1.

Method 3 is where I use strings. An empty grid would have a string gsFree = "1111....1111" (225) of these. Red checker pieces would be something like gsRed=" 0101010101010...00000" (225).

Functions and subs would pass these strings around, extract specific bits using MID( ) and CINT( ) to perform bit logic. Multiplying a string gsRed by 2 would be done by truncating MSB of string and adding a "0" for LSB.

Now I am leading towards Method 3 because it seems to make sense in my head. I am not worried about processor speed.

Regards,
ChuckB

Chuck
You are very eager to accomplish something, I can sense that now, and that is just great, but before you start to rush into any direction you have to ask yourself some questions, I'm not going to answer them, as I don't know what exactly you have in mind, and I don't want to ruin your enthusiasm either.

You are targetting to teach your bot mathematical concepts, there is already mathematical models for all our needs
1. Why would your bot need to learn them?
2. What would it be able to accomplish in your favour?
3. Why are there mathematical concepts?

I could answer on the third one of course, but I want to hear your answer first.

Chuck,

Love the idea of assigning 'boredom' and 'desperation' variables. Could also be neat if the AIP would accuse you of cheating (if it thought you claimed a win invalidly, for instance....)

All adds to the 'illusion'.

I don't think what most people consider to be 'true AI' will ever emerge from programs that seek to emulate human intelligence. We just don't understand our intelligence well enough to replicate it. I think we are far more likely to achieve MI (machine intelligence) through some (as yet) unknown process of allowing machines to self modify their behaviour - effectively reproduce and evolve.

well my simplistic view on AI:
I think humans may program AIs that will act intelligent (the same way we cosider humans to act intelligent), but only within their little system. Now take that to humans: Don't we just act intelligent in our 'little' system? Of course our system (the universe I guess) seems extremely complex to us. Don't you think there are 'things' (very vaguely I know) that might be way out of our intelilgence, maybe so impossible for us to understand that we don't even realize them.
So why shouldn't we archieve AI that is intelligent enough for what we are doing? I doubt there is a need in recreating human intelligence, don't we have enough of those?

I just think that as soon as you mention 'real AI' a lot of people immediately think Terminator , Lt. Cmdr Data or I, Robot :D

I just think that as soon as you mention 'real AI' a lot of people immediately think Terminator , Lt. Cmdr Data or I, Robot :D

kedaman I think Chuck just wants to do this more or less to see if he can - isn't that reason enough?

Quote:

---

Originally posted by kedaman

You are targetting to teach your bot mathematical concepts, there is already mathematical models for all our needs
1. Why would your bot need to learn them?
2. What would it be able to accomplish in your favour?
3. Why are there mathematical concepts?

---

1. Easy. Proof Of Concept. Simple mathematics (such as addition and multiplication) are fairly easy to understand, and, since they are a logical system, should be fairly easy to implement as a test. Once that is acheived, the method could be expanded to just about anything else.

2: Same as above. For the test case, there would be no "practical" value.

3: Eh?

Z.

Judd, sure he can, I am just very uncertain what exactly he wants to accomplish.
Misantrop, the universe you live in is full of tasks that you would consider boring, tedious or dangerous, AI is meant to perform these tasks for you, and some of them are very complicated and require sophisticated software.

Hi Everyone,

K, my personal objectives for my AIP gaming program is to design a program that can learn to play most games that a child of 7 to 10 can create. This will include games like TTT and chess.

Z speaks correctly about proof of concept. I am particularly interested in the time it will take for the program to go brain dead...5 mS, 2Sec, 14Min. I expect a flash of brillance and then a rapid darkhole...measured in a few seconds at most (and that's only because we use DoEvents) ;-)

J is absolutely right...I just want to do it to see if I can...and then if it yields surprisingly useful results I see potential applications for AIP game software.

M refers to our system the universe. My initial plan was to develop a natural language program (NLP) that could chat with my children. Sort of a Turing test for kids. However, I had to define a system or universe that both program and kids could relate to.

This led to gaming. Spend a day with an active 7 year-old and you will know what I mean. There world is wrapped around play.

Now it is conceivable to me that gaming actually introduces people to complex concepts in a natural way without all of the mathematics and lengthy explanations. I see that our model of the world around us can be expressed in terms of gaming. Therefore, AIP that chats and plays games with kids actually possesses the foundation in which we can build a more advanced NLP capability.

J likes the idea of emotion type variables. So do I. They lend to producing random acts of behavior, sudden changes in behavior and sometimes results in lucky decisions.

K, regarding rushing into this...its been five years on paper and in my head...its taken only 6 weeks on this forum for me to actually write what I am thinking (stage 1) in functional code. Without everyones comments and encouragement I may have become discouraged the 3rd time I wrote the AI engine for AIP. Thanks guys!

I am still looking forward to feedback on use of a super long integer using a string to simulate it for the purpose of using bit manipulations to simplify my process. Any experiences are also welcome.

Z, have you had a chance to consider what I said about the experiment above using my wife and also the possibility of a Calculator like panel in which buttons seemingly appear magically as the program discovers new concepts? I am curious to learn more about what you are thinking in this matter.

FINAL NOTE

AI has been a hobby for me...mostly reading lots of books. Generally the persons interested in AI are mathematicians and computer scientists. AI doesn't seem to be a good hobby like fishing or golf for a dumb guy like me. But how do I love it...

Regards,
ChuckB

Chuck this might be 'ahead of the game' (pun intended), but:

My fiancee was doing word searches in bed last night (hmmm...that sounds like I'm denigrating myself ;) ;) ) and I got to thinking abouut other types of pattern matching games that might work with AIPs. Word searches are one - but they don't really have a winner - just a winning condition.

But what about Solitaire (not the card game, the game with marbles)? Again, it's a 'winning condition' game, but I think an AIP could play itself and quickly develop a winning combination of moves (something I never managed with Solitarie!).

I also thought of a couple of other games I used to play as a kid, but I don't know if they have official names.

One was 'Squares' - where you start with a grid of dots and take turns to draw lines. If you complete a square, you claim it and get another go. The person with the most squares when the grid is full is the winner.

Another game was played on a symmetrical board (any shape, but must have an axis of symmetry). Players pieces (I don't remember how many, but the number isn't important) start in identical formations at either side of the board. They take turns to move, one piece at a time, the objective being to be the first to swap your formation to the opposite side of the board. As in checkers, you can only move forward. The game can be modified by allowing/disallowing 'jumps' over one of your opponents pieces(as in checkers).

Then I got to Nine-Men's Morris, Othello etc. etc.

I have a 'medieval' chess game at home too, that is simpler than traditional chess, but still a great strategy game.

Then there is 'Chinese Chess'.

I guess the list is very long....

Z & C

I took the time and read trough most of the posts around here, very quickly but still enough to grasp whats going on. The part which caught my interest was this idea of creating marcos:

Quote:

---

Maybe we can construct a simple program experiment where we define some AI, a goal, some key functions, program self-macro creation ability and see what happens? An example would be to write a program that can add and subtract. Give it ability to write macros...give it comprehension of summing. See if it can learn on its own to create a multiplication macro function on its own...and then use this to discover division, averages, etc.

---

My first question, why would it need to create macros in the first place? The answer i was expecting was to reduce complexity (the amount of operators in the expression). But to counter this, it would just add an extra cost for the reduction, it would still need to do every single operation including the reduction. In practice these concepts have no meaning to the process, just a matter of design. I'm sure this isn't what you were thinking of, so correct me, anyways, i will come back to this at the answer of my third question.

My second question, what would it benefit you if the program invented these concepts. Naturally if it does perform more efficient (which it doesn't) it would benefit you, so I wanted to know if it was in particular interest that you could have use of these concepts.

Third, why do we have mathematical concepts at all? This is the question to draw parallells between how human thinks and software processes information. As humans we have easier to remember well structured information, especially if we can use a couple of basic rules to derive a huge amount of data from, this way of storing information is called abstraction and is the reason why we are so much more flexible than most software, which are only equipped with minimal abstraction facilities. AI, especially with benefit from functional languages especially designed to deal with abstraction, makes it possible to derive similarly large amount of data from a small set of rules. Mathematical modelling is the art of using concepts to express symbolically any form of information processing in similar fashion, and we invented them to help us predict events and thus also achieving and preventing events. If an AI bot would need these, it would be to preserve memory usage.

I'm sorry if I don't really sound encouraging or enthusiastic, but I would be interested if you were going to have it learning mathematical concepts on its own. I'm not such an AI enthusiast like most others here are, but programming is a tedious task, and my goal is similar to AI but using the information we already have to accomplish software without human interference.

Hi K,

I have not read your postings as discouragement. You do talk over my head so it has taken a few days and several readings of your comments to understand what you are saying. Not your fault...I've taught myself to program without formal programming instruction so I missed out on a lot.. :-)

I see you are clearly focused upon the side of efficiency and the use of AI to reduce the tedium of programming. If AI isn't efficient...or doesn't simplify the processing capability of the application...then you ask the question "What's the point?"

I am NOT interested in efficiency. Maybe that disqualifies me from being an AI purist...not sure. But AI in my opinion must benefit humans directly. If it satisfies this requirement then I consider it legitimate.

Consider this from the "human" side of the problem...which I equate as being far more important in most cases.

Question 1
=======
You said the 'macro adds no meaning' to the process. Now this is not right from my frame of reference. Say you use an AI calculator with a +, - and = key. You are limited by those operations for solving math problems. Suppose your AI program very inefficiently and redundantly derives another operation such as x (multiplication). A button pops up on the AI calculator and Voila!...you have now increased your computational power as a human. From the human's perspective, we have just achieved a significant degree of meaning with respect to our ability. Nevermind the wasted CPU computational cycles (technology will soon correct that problem).

Question 2
========
Benefit to the above scenario is straightforward. If you have taken an advanced math exam in college then you will understand that a scientific graphing calculator is far superior to a calculator that only adds and subtracts (although grandfather tells stories of how he used to use an abacus to solve complex physics problems. :-)

Question 3
========
Abstraction? I consider AIP v0.2a to be an excellent experiment in abstraction. My 7 year-old wrote a computer program to play tic-tac-toe with AIP just by using what humans do best...playing. How abstract can I get K?

You say "...my(your) goal is similar to AI but using the information we already have to accomplish software without human interference."

I agree. In the same way that a 'human' must program a computer, humans can 'teach' the computer. Both methods then allow the computer to operate on its own with existing information that does not require human intervention. Macro development as Z and I have discussed definitely satisfies that requirement.

Isn't is cool that we are so much alike? ;-) Keep writing...don't be discouraged if I don't accept traditional AI thinking.

The more and more I study apparently complex things such as colonies of bugs, human DNA decoding, creation of life...I realize that it really isn't so complex as we make it to be. It is all actually much simpler...the problem is we just like complex solutions and algoritms. ;-)

Regards,
ChuckB

Hello C, I'm definitely not trying to discourage you, but I'm rather trying to figure out if it would interest me and also help you if i can.

I have to try and explain more i know, I always want things concise but whats the use if I'm not understood, so if I'm not making any sense let me know :) Let us start from the beginning again and make sure what we don't go to fast.

The vital step you are making is to introduce macros, in order to present new operations to the user, for instance multication, based on addition. Your software can't do it without two things, the need to invent it, as well as inductive reasoning.

A very primitive human, who doesn't know much math at all is introduced to a problem, this person only knows what weight and length is and how to measure length. He can carry certain weight. Now until now he has been carrying blocks of various length, but with fixed height and thickness. By measuing length he can decide if he can carry the piece or not.
The problem is that some pieces are too heavy for him, even though they are shorter than the limit. He figures out that the block height is varying and that this block has more height than most of the blocks. After some inductive reasoning he decides that the weight is calculated as follows:

weight = length + height

well after a while he discovers that this doesn't work either and he has to do more thinking. He finally decides to cut up the blocks into smaller cube blocks all of the same size. and the blocks are arranged 3x4. He understands why the formula doesn't work, and decides he arranges the block into 4 groups of 3, height+height+height+height, and as he can see that there are 4 groups and the length is 4 he decides that there must be a connection. He now knows that he has to count height, length times to calculate the weight, and he decides to call it length times height.

All these decisions are derived from various experiences with all kinds of situations, that doesn't ever occurr to computer program. Even trough testing empirically with incredible processing power, there's no ground to give hints to make these decisions, it took the evolution billions of years to evolve our brains and accumulate the empirical test results in our genes.

What if we just told the AI what to do, F=w*l*h*rho*g (the weight is width times length times height times density times gravity) and exactly how you would calculate with floating point precision, and it could solve all problems of this type. If we transfered all human knowledge over to AI, it could at some point also be able to do empirical guesses as we do.

I can link directly to this thread, but it doesn't appear in the 'Games and Graphics Programming' section.

Tum te tum.

Chuck, I've been trying to figure out if there is a 'perfect' knowledge array for TTT wrt the AIP. Is there one?

I wanted to start with a 'perfect' AIP and have it teach another until they lock onto the tied game, then take the 2nd generation AIP and teach the 3rd generation until tie occurs etc.

Will they all get to the smae tied game?

Hi Judd,
Perfect knowledge array for TTT. The game is pretty limited. There are 8 winning patterns. The center cell is the best because it is related to 4 patterns, corners are next with 3 patterns and the remaining cells intersect only 2 patterns each.

All it takes is for the teacher to play 4 winning patterns before AIP is fully educated. (2 diagonal wins and 1 vertical and 1 horizontal win). It derives the remaining 4 patterns due to its hardcoded ability to recognize patterns (like humans).

AIP stores winning patterns in the file as follows:

ABS(1,1);(2,1);(3,1)=3 'left vertical pattern
ABS(1,2);(2,2);(3,2)=4 'center vertical pattern
ABS(1,3);(2,3);(3,3)=5 'right vertical pattern

As the game progresses, AIP compares the current state of the game with winning patterns and free cells to produce something that looks a bit like this:

100 000 100 = 2
000 010 000 = 1
001 001 001 = 3
=========
101 011 101

The HSum of 2,1 and 3 tells AIP how many moves are required to produce this win. '1' is the answer which corresponds to center vertical column. If the HSums are equal, then AIP examines the VSum. Anything with a '0' is not considered. Usually the highest VSum would be selected.

AIP v0.2 does not use the history data at the top of 3,4 or 5. This means that in AIP's experience, the right vertical column had the better chance of winning. This would be valuable against a human because as humans we have our predicatable patterns and ways of doing things. A computer program would not have that bias and therefore historical data may not be so important.

Now, v0.3 will have these two improvements: 1) use historical data to force AIP to decide in the event that all VSums are the same value or all HSums are the same value or 2) after playing the same moves for 2 consecutive games, it will randomly pick a different move in order to break out of the mold.

Therefore, experience (playing lots of games) will lead AIP to learn more winning patterns and to collect statistical data on what patterns seem to win the most.

With these improvements we will see different strategies emerge but eventually I believe the AIP teacher and AIP student will continue to tie.

Judd, here is the problem that need figuring out. I am looking ahead at v0.4 that will hopefully play checkers/chess. Teaching a program to learn moves will be relatively easy if I model it after a human being. The difficult part, which I have not figured out, is how can experience (playing lots of games) make AIP smarter and smarter...eventually playing at a level as a grandmaster (...now I am in fantasy land...just humor me)? You see, this generic algorithm must be hardcoded in advance. TTT was easy...this whoops up on me.

Now I am thinking this...

perhaps AIP will begin stringing moves together to complete a goal. For example, it takes two moves of a knight to capture a piece that is normally out of range...or a piece can be made to block, capture, sacrifice or trap.

Kedaman, I haven't forgotten your last post. I'll respond shortly.

Regards,
ChuckB