Artificial, but Intelligent? Part 2

I just finished reading Practical Game AI Programming: Unleash the power of Artificial Intelligence to your game. My conclusion is there is a lot less to this book than meets the eye.

For someone thinking of purchasing this book, it would be difficult to weigh that decision before committing. The above link to Amazon has (as of this writing) no reviews. I’ve not found any other, independent evaluations of this work. Perhaps you could make a decision simply by studying the synopsis of this book before you buy it. Having done that, it is possible that you’d be prepared for what it offered. Having read the book, and then going back and reading the Amazon summary (which is taken from the publisher’s website), I find that it more or less describes the book’s content. In my case, I picked this book up as part of a Humble Book Bundle, so it was something of an impulse buy. I didn’t dig too hard into the description and instead worked my way through the chapters with my only expectations being based on the title.

Even applying the highest level of pre-purchase scrutiny only gets you so far. The description may indicate that the subject matter is of interest, but it is still a marketing pitch. It gives you no idea of the quality of either the information or the presentation. Furthermore, I think someone got a little carried away with their marketing hype. The description also tosses out some technical terms (e.g. rete algorithm, forward chaining, pruning strategies) perhaps meant to dazzle the potential buyer with AI jargon. The problem is, these terms don’t even appear in the book, much less get demonstrated as a foundation for game programing. I feel that no matter how much upfront research you did before you bought, you’d come away feeling you got less than you bargained for.

What this book is not is a exploration of artificial intelligence as I have discussed that term previously on this website. This is not about machine learning or generic decision-making algorithms or (despite the buzz words) rule-engines. The book mentions applications like Chess only in passing. Instead, the term “AI” is used as a gamer might. It discusses a few tricks that a game programmer can use to make the supposedly-intelligent entities within a game appear to have, well, intelligence when encountered by the player.

The topic that it does cover does, in fact, have some merit. The focus is mostly on simple algorithms and minimal code required to create the impression of intelligent characters within a game. Some of the topics I found genuinely enlightening. The overarching emphasis on simplicity is also something makes sense for programmers to aspire to. There is no need to program a character to have a complex motivation if you can, with only a few lines of code, program him to appear to have such complex motivation. It is just that I’m not sure that these lessons qualify as “unleashing the power of Artificial Intelligence” by anyone’s definition.

But even before I got that far, my impression started off very bad. The writing in this book is rather poor, in terms of grammar, word usage, and content. In some cases, misused words become so jarring as to make it difficult to read through a page. Elsewhere, there will be several absolutely meaningless sentences strung together, perhaps acknowledging that a broader context is required but not knowing how to express it. At first, I didn’t think I was going to get very far into the book. After a chapter or so, however, reading became easier. Part of it may be my getting used to the “style,” if one can call it that. Part of it may also be that there is more “reaching” in the introductory and concluding sections but less when writing about concrete points.

I can’t say for sure but it is my guess, based on reading through the book, that the author does not use English as his primary language. I sometimes wondered if the text was written first in another language and then translated (or mistranslated, as the case may be) into English. Beyond that, the book also does not seem to have benefited from the work of a competent editor.

The structure of the chapters, for the most part, follows a pattern. A concept is introduced by showcasing some “classic” games with the desired behavior. Then some discussion about the principle is followed by coding example, almost always in Unity‘s C# development environment. This is often accompanied by screenshots of Unity’s graphics, either in development mode or in run-time. Most of the chapters, however, feel “padded.” Screenshots are sometimes repetitious. Presentation of the code is done incrementally, with each new addition requiring the re-printing of all of the sample code shown so far along with the new line or lines added in. By the end of the chapter, adding a concept might consist of two explanatory sentences, 3 screenshots, and two pages of sample code, 90% of which is identical to the sample code several pages earlier in the book. This is not efficient and I don’t think it is useful. It does drive the page count way up.

I want to offer a caveat for my review. This is the first book I’ve read from this publisher. When reading about some of their other titles, it was explained that the books come with sample source code. If you buy the book directly from the publisher’s website (which I did not), the sample code is supposed to be downloaded along with the book text. If you buy from a third party, they provide a way to register your purchase on the publisher’s site to get access to the downloads. I did not try this. If this book does have downloaded samples that can be loaded into Unity, and those samples are well-done, that has a potential for adding significant value over the book on its own.

Back to the chapters. When I start going through the chapters, again it feels like there is some “padding” going on to make the subject matter seem more extensive than it is. The book starts with two chapters on Finite State Machines FSM and how that logic can be used to drive an “AI” character’s reactions to the player. Then the book takes a detour into Unity’s support for a Finite State Machine implementation of animations, which has its own chapter. This is most irrelevant to the subject of game AI and also, likely, of little value if you’re not using Unity.

After the animation chapter, we head back into the AI world with a discussion of the A*, and the Theta* variant thereof, pathfinding algorithm. This discussion is accompanied by a manual optimization solution of a simple square-grid based 2D environment, describing each calculation and illustrating each step. I do appreciate the concrete example of the algorithm in action. Many explanations of this topic I’ve found on-line simply show code or pseudo-code and leave it to the “student” to figure it all out. In this case, I think he managed to drive the page count up by and order of magnitude over what would have been sufficient to explain it clearly.

The final chapters show how Unity’s colliders and raycasting can be used to implement both collision avoidance and vision/detection systems. These are two very similar problems involving reacting to other objects in the environment that, themselves, can move around. As I said earlier, there are some useful concepts here, particularly in emphasizing a “keep it simple” design philosophy. If you can use configurable attributes on your development tool’s existing physics system to do something, that’s much preferable to generating your own code base. That goes double if the perception for the end user is indistinguishable, one method from the other. However, I also get the feeling that I’m just being shown some pictures of simple Unity capabilities, rather than “unleashing the power of AI” in any meaningful sense.

A few year’s back, I was trying to solve a similar problem, but trying to be predictive about the intent of the other object. For example, if I want to plot an intercept vector to a moving target but that target is not, itself, moving at a constant rate or direction, I need a good bit more math than the raycasting and colliders provide out of the box. Given the promise of this book’s subject matter, that might be a problem I’d expect to find, perhaps in the next chapter.

Alas, after discussing the problem of visual detection involving both direction and obstacles, the book calls an end to its journey. With the exception of the A* algorithm, the AI solutions consist almost entirely of Unity 3D geometry calls.

Although the book claims to be written in a way such that each chapter can be applied to a wide range of games, I feel like it narrows its focus as it progresses. The targeted game is, and I struggle with how to describe it so I’ll just pick an example, the heirs to the DOOM legacy. By this, I mean games where the player progresses through a series of “levels” in order to complete the game. What the player encounters through those levels is imagined and created by the designer so as to construct the story of the game. The term AI, then, distinguishes between different kinds of encounters, at least as far as the player perceives them. For example, the player might find herself rushing across a bridge, which starts to collapse when they reach the middle. This requires no “AI.” There is simple programmed in that, when the player reaches a certain point on the bridge, call the “collapseBridge” routine. If she makes it past the bridge and into the next chamber, where there are a bunch of gremlins that want to do her in, the player starts considering the “AI” of those gremlins. Do they react to what the player does, adopting different tactics depending on her tactics? If so, she might praise the “AI.” By the books end, the focus is entirely on awareness of and reaction between mobile elements of a game which, by defining the problem as such, is the subset of games in this category.

My harping on the narrow focus of this book goes to the determination of its value. If this book were free or very low cost, you would have to decide whether the poor use of English and the style detract from whatever useful information is presented. The problem with that is the price this book asks. The hardcopy (paperback) of the book is $50.00. The ebook is $31.19 on Amazon, discounted to $28 if you buy directly from the publisher’s site. All of those seem like a lot of money, per my budget. Now, my own price I figure to have been $7. I bought the $8 bundle package over the $1 package purely based on interest in this title. This is the first book in that set I’ve read, so if some of the others are good, I might consider the cost to be even lower. Still, even at $5, I feel like I’ve been cheated a bit by the content of this book.

The bundle contained other books from this same publisher, so I’ll plan to read at least one other before drawing any conclusions about their whole library. Assuming that the quality of this book is, in fact, an outlier, this is still a risk to the publisher’s reputation. When one of your books is overpriced and oversold, the cautious buyer should assume that they are all overpriced and oversold. Looking at the publisher’s site, this book has nothing but positive reviews. It’s really a blemish on the publisher as a whole.

Although I won’t go so far as to say “I wish I hadn’t wasted the time I spent reading this,” I can’t imagine any purchaser for whom this title would be worth the money.

The Pride and Disgrace

Handful of Senators don’t pass legislation.

On August 10th, 1964, the Gulf of Tonkin Resolution was enacted. It was the quick response from America’s politicians to the attack upon a U.S. naval vessel off the coast of Vietnam by torpedo boats from the communist regime of North Vietnam. That attack had occurred on August 2nd with an additional “incident” taking place on August 4th. The resolution (in part) authorized the President “to take all necessary steps, including the use of armed force, to assist any member or protocol state of the Southeast Asia Collective Defense Treaty requesting assistance in defense of its freedom.” This would come to justify the deployment of U.S. troops to directly engage the enemies of South Vietnam.

Before this time, South Vietnam was fighting to eliminate a communist insurgency driven by remnants of the Việt Minh. The communist guerillas had agreed to resettle in the North as part of the Geneva Agreement which ended France’s war in Vietnam in 1954. The United States saw their continued attacks on the government of South Vietnam as a violation of that peace. In particular, their support obtained from across national borders was considered to be a part of a strategic plan by the Soviet Union and China to spread communism throughout all of the countries of Southeast Asia.

Even before the withdrawal of France, the United States had supported the anti-communist fight with money, matériel, and military personnel (in the form of advisors). After the French exit, a steady increase in commitment from the U.S was evident. Nevertheless, the signing of the Gulf of Tonkin resolution marks a milestone in American’s involvement making it arguable the start of the U.S. War in Vietnam. Although it would take almost another year for U.S. forces to become clearly engaged, Presidents Johnson’s reaction to the Gulf of Tonkin incidents seems to have set the course inevitably toward that end.

I’m sittin’ here.
Just contemplatin’.
I can’t twist the truth.
It knows no regulation.

The anniversary and the nature of it got me to thinking about how one might portray the Vietnam War as a game. Given the context, I’m thinking purely along the lines of the game focused at the strategic level, taking into account the political and international considerations that drove the course of the conflict.

From a high-level perspective, one might divide America’s war in Vietnam into 4 distinct phases. In the first, the U.S. supported Vietnam’s government with financial and military aid, and with its advisors. While U.S. soldiers were, in fact, engaging in combat and being killed, it wasn’t as part of American combat units, allowing the U.S. to convince itself that this was a conflict purely internal to South Vietnam. Through the presidencies of Eisenhower, Kennedy, and Johnson, the amount of aid to, and the number of Americans in, Vietnam increased. However, the big change, and the transition of the second phase, can be located after the passage of the Gulf of Tonkin Resolution and the subsequent deployment of the U.S. Marines directly and as a unit.

At first, U.S. direct involvement carried with it a measure of popular support and was, from a purely military standpoint, overwhelmingly successful. Johnson and the military were wary of pushing that involvement in ways that would turn the public opinion against them. The U.S. feared casualties as well signs of escalation that be interpreted as increasing military commitment (for example, extending the service time for draftees beyond twelve months), but in general this was a period of an increasing U.S. buildup and, generally, successful operations. Nevertheless, progress in the war defied a clear path toward resolution.

The third phase is probably delineated by the 1968 Tet offensive. While still, ultimately, a military success from the U.S. standpoint, the imagery of Viet Cong forces encroaching on what were assumed by all to be U.S.-controlled cities turned opinion inexorably against continuing engagement in Vietnam. The next phase, then, was what Nixon called “Vietnamization,” the draw-down of American direct involvement to be replaced with support for the Army of the Republic of Vietnam (ARVN). Support was again in the form of money, equipment, and training as well as combat support. For example, a transition to operations where ARVN ground units would be backed by U.S. air power.

The final phase is where that withdrawl is complete, or at least getting close to that point. Where joint operations were no longer in the cards. Clearly this phase would describe the post-Paris accords situation, after Nixon’s resignation, as well as encompassing the final North Vietnamese operation that rolled up South Vietnam and Saigon.

From a gaming perspective, and a strategic-level gaming perspective at that, the question becomes what decisions are there for a player to make within these phases and, perhaps more importantly, what decisions would prompt the transition from one phase to another.

The decision to initially deploy U.S. troops, made by Johnson in early 1965, seems to have been largely driven by events. Having Johnson as president was probably a strong precondition. Although he ran against Goldwater on a “peace” platform, the fact that he saw his legacy as being tied into domestic policy probably set up the preconditions for escalation. A focus on Vietnam was never to be part of his legacy, but given the various triggers in late 1964 and early 1965, his desire to avoid a loss to communism in Vietnam propelled his decision to commit ground troops. You might say his desire to keep it from being a big deal resulted in it being a big deal.

Where this all seems to point is that any strategic Vietnam game beginning much before Spring of 1965 must restrict the player from making the most interesting decision; if and when to commit U.S. ground troops and launch into an “American” war.

Amusingly, if you subscribe to the right set of conspiracy theories, the pivotal events might really be under control of a grand-strategic player after all. Could it be that the real driver behind Kennedy’s assassination was to put a President in office who would be willing to escalate in Vietnam? Was the deployment of the USS Maddox on the DESOTO electronic warfare mission meant to provoke a North Vietnamese response? How about the siting of aviation units in Vietnam at places like Camp Holloway, which would become targets for the Viet Cong? Where actual aggression by the North wasn’t sensational enough, were details fabricated? This rather far-flung theorizing would not only make the resulting game that much harder to swallow, but it is also difficult to see how any fully-engineered attempt to insert American into Vietnam could have moved up the timetable.

So it would only make sense to start our game with our second phase, which must come after our Gulf of Tonkin incident and the 1968 presidential election, at a minimum.

The remaining game will still be an unconventional one, although we do have some nice examples of how it could be done from over the years. Essentially, the U.S. will always be able to draw upon more military power and, ultimately, sufficient military power to prevail in an particular engagement. Yet while it is possible, though insufficient planning, to achieve a military loss as the U.S. it is probably not going to be possible to achieve a military victory. On the U.S. side, the key parameters are going to be some combination of resources and “war weariness.”

Our Vietnam game would rule out, either explicitly or implicitly, a maximal commitment to victory by the United States. American planners considered options such as unfettered access to Laos and Cambodian, an outright invasion of North Korea, or even tactical nuclear weapons. The combination of deteriorating domestic support and the specter of overt Chinese and Soviet intervention would seem to be a large enough deterrent to prevent exercise of these options. This is one of the reasons that rules (those that I’ve come across, any way) simply forbid, for example, crossing units into North Vietnam.

The other reason is one of scope. If a ground invasion of North Vietnam is on the table, then the map needs to include all the potential battlefields in the North in addition to the actual battlefields of the South. Likewise extended areas within Cambodia and Laos need to be available to the player. Continuing on, if U.S. ground forces are going to be straying that close to North Vietnam’s northern border, might it not be necessary to include China in as well? Perhaps having learned a lesson from Korea, our player would react to Chinese direct intervention by taking the fight onto Chinese sovereign territory. It doesn’t take long before we have to consider adding Germany, Korea, Cuba, and any other hotspot of the time as a potential spillover for escalation in Vietnam. Besides the problem of the game expanding without limitations, we have another design concern. A Vietnam game narrative adhering closely to the historical path has the advantage of actual battles, strategies, and events on which to model itself. If all the forces of NATO, the Warsaw Pact, and China are fair game, we are now in the realm of pure speculation.

If for no other reason than to maintain the sanity of the designer, it seems that rules which quickly push the U.S. into its historical de-escalation policy is the right way to tie off such a game on the other end. I will save the consideration of how that might work to another time.

Artificial, Yes, but Intelligent?

Keep your eyes on the road, your hands upon the wheel.

When I was in college, only one of my roommates had a car. The first time it snowed, he expounded upon the virtues of finding an empty and slippery parking lot and purposely putting your car into spins.  “The best thing about a snow storm,” he said. At the time I thought he was a little crazy. Later, when I had the chance to try it, I came to see it his way. Not only is it much fun to slip and slide (without the risk of actually hitting anything), but getting used to how the car feels when the back end slips away is the first step in learning how to fix it if it happens when it matters.

Recently, I found myself in an empty, ice-covered parking lot and, remembering the primary virtue of a winter storm, I hit the gas and yanked on the wheel… but I didn’t slide. Instead, I encountered a bunch of beeping and flashing as the electronic stability control system on my newish vehicle kicked in. What a disappointment it was. It also got me thinkin’.

For a younger driver who will almost never encounter a loss-of-traction slip condition, how do they learn how to recover from a slide or a spin once it starts? Back in the dark ages, when I was learning to drive, most cars were rear-wheel-drive with a big, heavy engine in the front. It was impossible not to slide around a little when driving in a snow storm. It was almost a prerequisite to going out into the weather to know all the tricks of slippery driving conditions. Downshifting (or using those number gears on your automatic transmission), engine breaking, and counter steering were all part of getting from A to B. As a result*, when an unexpectedly slippery road surprises me, I instinctively take my foot off the brakes/gas and counter-steer without having to consciously remember the actual lessons. So does a car that prevents sliding 95% of the time result in a net increase in safety, even though it probably makes that other 5% worse? It’s not immediately obvious that it does.

On the Road

I was reminded of the whole experience a month or so ago when I read about the second self-driving car fatality. Both crashes happened within a week or so of each other in Western states; the first in Arizona and the second in California. In the second crash, Tesla’s semi-autonomous driving function was in fact engaged at the time of the crash and the drivers hands were not on the wheel six seconds prior. Additional details do not seem to be available from media reports, so the actual how and why must remain the subject of speculation. In the first, however, the media has engaged in the speculation for us. In Arizona, it was an Uber vehicle (a Volvo in this case) that was involved and the fatality was not the driver. The media has also reported quite a lot that went wrong. The pedestrian who was struck and killed was jaywalking, which certainly is a major factor in her resulting death. Walking out in front of a car at night is never a safe thing to do, whether or not that car self-driving. Secondly, video was released showing the driver was looking at something below the dashboard level immediately before the crash, and thus was not aware of the danger until the accident occurred. The self-driving system itself did not seem to take any evasive action.

Predictably, the Arizona state government responded by halting the Uber self-driving car program. More on that further down, but first look at the driver’s distraction.

After the video showing such was released, media attention focused on the distracted-driving angle of the crash. It also brought up the background of the driver, who had a number of violations behind him. Certainly the issue of electronics and technology detracting from safe driving is a hot topic and something, unlike self-driving Uber vehicles, that most of us encounter in our everyday lives. But I wonder if this exposes a fundamental flaw in the self-driving technology?

It’s not exactly analogous to my snow situation above, but I think the core question is the same. The current implementation of the self-driving car technology augments the human driver rather then replaces him or her. In doing so, however, it also removes some of the responsibility from the driver as well as making him more complacent about the dangers that he may be about to encounter. The more that the car does for the driver, the greater the risk that the driver will allow his attention to wander rather that stay focused, on the assumption that the autonomous system has him covered. In the longer term, are there aspects of driving that the driver will not only stop paying attention to, but lose the ability to manage in the way a driver of a non-automated car once did?

Naturally, all of this can be designed into the self-driving system itself. Even if a car is capable of, essentially, driving itself over a long stretch of a highway, it could be designed to engage the driver every so many seconds. Essentially requiring unnecessary input from the operator can be used to make sure she is ready to actively control the car if needed. I note that we aren’t breaking new ground here. A modern aircraft can virtually fly itself, and yet some part of the design (plus operational procedures) are surely in place to make sure that the pilots are ready when needed.

As I said, the governmental response has been to halt the program. In general, it will be the governmental response the will be the biggest hurdle for self-driving car technology.

In the specific case of Arizona, I’m not actually trying to second guess their decision. Presumably, they set up a legal framework for the testing of self-driving technology on the public roadways. If the accident in question exceeded any parameters of that legal framework, then the proper response would be to suspend the testing program. On the other hand, it may be that the testing framework had no contingencies built into it, in which case any injuries or fatalities would have to be evaluated as they happen. If so, a reactionary legal response may not be productive.

I think, going forward, there is going to be a political expectation that self-driving technology should be flawless. Or, at least, perfect enough that it will never cause a fatality. Never mind that there are 30-40,000 motor vehicle deaths per year in the United States and over a million per year world wide. It won’t be enough that an autonomous vehicle is safer than than a non-autonomous vehicle; it will have to be orders-of-magnitude safer. Take, as an example, passenger airline travel. Despite a rate that is probably about 10X safer for aircraft over cars, the regulatory environment for aircraft is much more stringent. Take away the “human” pilot (or driver) and I predict the requirements for safety will be much higher than for aviation.

Where I’m headed in all this is, I suppose, to answer the question about when we will see self driving cars. It is tempting to see that as a technological question – when will the technology be mature enough to be sold to consumers? But it is more than that.

I recall see somewhere an example of “artificial intelligence” for a vehicle system. The example was of a system that detected a ball rolling across the street being a trigger for logic that anticipates there might be a child chasing that ball. A good example of an important problem to solve before putting an autonomous car onto a residential street. Otherwise, one child run down while he was chasing his ball might be enough for a regulatory shutdown. But how about the other side of that coin? What happens the first time a car swerves to avoid a non-existent child and hits an entirely-existent parked car? Might that cause a regulatory shutdown too?

Is regulatory shutdown inevitable?

Robo-Soldiers

At roughly the same time that the self-driving car fatalities were in the news, there was another announcement, even more closely related to my previous post. Video-game developer EA posted a video showing the results of a multi-disciplinary effort to train a AI player for their Battlefield 1 game (which, despite the name is actually the fifth version of the Battlefield series). The narrative for this demo is similar to that of Google’s (DeepMind) chess program. The training was created, as the marketing pitch says, “from scratch using only trial and error.” Without viewing it, it would seem to run counter to my previous conclusions, when I figured that the supposed generic, self-taught AI was perhaps considerably less than it appeared.

Under closer examination, however, even the minute-and-a-half of demo video does not quite measure up to the headline hype, the assertion that neural nets have learned to play Battlefield, essentially, on their own. The video explains that the training methods involves manually placing rewards throughout the map to try to direct the behavior of the agent-controlled soldiers.

The time frame for a project like this one would seem to preclude them being directly inspired by DeepMind’s published results for chess. Indeed, the EA Technical Director explains that it was earlier DeepMind work with Atari games that first motivated them to apply the technology to Battlefield. Whereas the chess example demonstrated ability to play chess at a world class level, the EA project demonstration merely shows that the AI agents grasp the basics of game play and not much more. The team’s near-term aspirations are limited; use of AI for quality testing is named as an expected benefit of this project. He does go so far as to speculate that a few years out, the technology might be able to compete with human players within certain parameters. Once again, a far cry from a self-learning intelligence poised to take over the world.

Even still, the video demonstration offers a disclaimer. “EA uses AI techniques for entertainment purposes only. The AI discussed in this presentation is designed for use within video games, and cannot operate in the real world.”

Sounds like they wanted to nip any AI overlord talk in the bud.

From what I’ve seen of the Battlefield information, it is results only. There is no discussion of the methods used to create training data sets and design the neural network. Also absent is any information on how much effort was put into constructing this system that can learn “on its own.” I have a strong sense that it was a massive undertaking, but no data to back that up. When that process becomes automated (or even part of the self-evolution of a deep neural network), so that one can quickly go from a data set to a trained network (quickly in developer time, as opposed to computing time), the promise of the “generic intelligence” could start to materialize.

So, no, I’m not made nervous that an artificial intelligence is learning how to fight small unit actions. On the other hand, I am surprised at how quickly techniques seem to be spreading. Pleasantly surprised, I should add.

While the DeepMind program isn’t open for inspection, some of the fundamental tools are publicly available. As of late 2015, the Google library TensorFlow is available in open source. As of February this year, Google is making available (still in beta, as far as I know) their Tensor Processing Unit (TPU) as a cloud service. Among the higher-profile uses of TensorFlow is the app DeepFake, which allows its users to swap faces in video. A demonstration compares the apps performance, using a standard desktop PC and about a half-an-hour’s training time to produce something comparable to Industrial Light and Magic’s spooky-looking Princess Leia reconstruction.

Meanwhile, Facebook also has a project inspired by DeepMind’s earlier Go neural network system. In a challenge to Google’s secrecy, the Facebook project has been made completely open source allowing for complete inspection and participation in its experiments. Facebook announced results, at the beginning of May, of a 14-0 record of their AI bot against top-ranked Go players.

Competition and massive-online participation is bound to move this technology forward very rapidly.

 

The future’s uncertain and the end is always near.

 

*To be sure, I learned a few of those lessons the hard way, but that’s a tale for another day.

ABC Easy as 42

Teacher’s gonna show you how to get an ‘A’

In 1989, IBM hired a team of programmers out of Carnegie Mellon University. As part of his graduate program, team leader Feng-hsiung Hsu (aka Crazy Bird) developed a system for computerized chess playing that the team called Deep Thought. Deep Thought, the (albeit fictional) original, was the computer created in Douglas Adam’s The Hitchhiker’s Guide to the Galaxy to compute the answer for Life, the Universe, and Everything. It was successful in determining the answer was “42,” although it remained unknown what the question was. CMU’s Deep Thought, less ambitiously, was a custom designed hardware-and-software solution for solving the problem of optimal chess playing.

Once at IBM, the project was renamed Deep Blue, with the “Blue” being a reference to IBM’s nickname of “Big Blue.”

On February 10th, 1996, Deep Blue won its first game against a chess World Champion, defeating Garry Kasparov. Kasparov would go on to win the match, but the inevitability of AI superiority was established.

Today, computer programs being able to defeat humans is no longer in question. While the game of chess may never be solved (à la checkers), it is understood that the best computer programs are superior players to the best human beings. Within the chess world, computer programs only make news for things like when top players may using programs to gain an unfair advantage in tournament play.

Nevertheless, a chess-playing computer was in the news late last year. Headlines reported that a chess playing algorithm based on neural networks, starting only from the rules of legal chess moves, in four hours created a program that could beat any human and nearly all top-ranked chess programs. The articles spread across the internet through various media outlets, each summary featuring their own set of distortions and simplifications. In particular, writers that had been pushing articles about the impending loss of jobs to AI and robots jumped on this as proof that the end had come. Fortunately, most linked to the original paper rather than trying to decipher the details.

Like most I found this to be pretty intriguing news. Unfortunately, I also happen to know a little (just a little, really) about neural networks, and didn’t even bother to read the whole paper before I started trying to figure out what had happened.

Some more background on this project. It was created at DeepMind, a subsidiary of Alphabet, Inc. This entity, formerly known simply as Google, reformed itself in the summer of 2015 with the new Google being one of many children of the Alphabet parent. Initial information suggested to me an attempt at creating one held company for each letter of the alphabet, but time has shown that isn’t their direction. As of today, while there are many letters still open, several have multiple entries. Oh well, it sounded more fun my way. While naming a company “Alphabet” seems a bit uninspired, there is a certain logic to removing the name Google from the parent entity. No longer does one have to wonder why an internet company is developing self-driving cars.

Google’s self driving car?

 

The last time the world had an Artificial Intelligence craze was in the 1980s into the early 1990s. Neural networks were one of the popular machine intelligence techniques of that time too. At first they seemed to offer the promise of a true intelligence; simply mimicking the structure of a biological brain could produce an ability to generalize intelligence, without people to craft that intelligence in code. It was a program that could essentially teach itself. The applications for such systems seemed boundless.

Unfortunately, the optimism was quickly quashed. Neural networks had a number of flaws. First, they required huge amounts of “training” data. Neural Nets work by finding relationships within data, but that source data has to be voluminous and it has to be suited to teaching the neural network. The inputs had to be properly chosen, so as to work well with the networks’ manipulation of that data and the data themselves had be properly representative of the space being modeled. Furthermore, significant preprocessing was required of the person organizing the training. Additional inputs would result in exponential increases in both the training data requirement and the amount of processing time to run through the training.

It is worthwhile to recall the computer power available to neural net programmers of that time. Even a high-end server of 35 years ago is probably put to shame by the Xbox plugged into your television. Furthermore, the Xbox is better suited to the problem. The mathematics capability of Graphical Processing Units (GPUs) is a more efficient design for solving these kinds of matrix problems. Just like Bitcoin mining, it is the GPU on a computer that is going to best be able to handle neural network training.

To illustrate, let me consider briefly a “typical” neural network application of the previous generation. One use is something called a “soft sensor.” Another innovation of that same time was the rapid expansion in capabilities of higher-level control systems for industrial processes. For example, some kind of factory wide system could collect real-time data (temperatures, pressures, motor speeds – whatever is important) and present them in an organized fashion to give an overview of plant performance and, in some cases, automate overall plant control. For many systems however, the full picture wasn’t always available in real time.

Let’s imagine the production of a product which has a specification limiting the amount of some impurity. Largely, we know what the right operating parameters of the system are (temperatures, pressures, etc) but to actually measure for impurities, we manually draw a sample, send it off to a lab for testing, and wait a day or two for the result. It would stand to reason that, in order to keep your product within spec, you must operate far enough away from the threshold that if it begins to drift, you would usually have time catch it before it goes out of spec. Not only does that mean you’re, most of the time, producing a product that exceeds specification (presumably at extra cost), but if the process ever moves faster than expect, you may have to trash a day’s worth of production created while you were waiting for lab results.

Enter the neural network and that soft sensor. We can create a database of the data that were collected in real time and correlate that data with the matching sample analyses that were available afterward. Then a neural network can be trained using the real-time measurements as input to produce an output predicting sample measurement. Assuming that the lab measurement is deducible from the on-line data, you now have in your automated control system (or even just as a presentation to the operators) a real time “measurement” of data that otherwise won’t be available until much later. Armed with that extra knowledge, you would expect to both cut operating costs (by operating tighter to specification) and prevent waste (by avoiding out of spec conditions before they happen).

That sounds very impressive, but I did use the word “assuming.” There were a lot factors that had to come together before determining that a particular problem was solvable with neural networks. Obviously, the result you are trying to predict has to, indeed, be predictable from the data that you have. What this meant in practice is that implementing neural networks was much bigger than just the software project. It often meant redesigning your system to, for example, collect data on aspects of your operation that were never necessary for control, but are necessary for the predictive functioning of the neural net. You also need lots and lots of data. Operations that collected data slowly or inconsistently might not be capable of providing a data set suitable for training. Another gotcha was that collecting data from a system in operation probably meant that said system was already being controlled. Therefore, a neural net could just as easily be learning how your control system works, rather than the underlying fundamentals of your process. In fact, if your control reactions were consistent, that might be a much easier thing for the neural net to learn that the more subtle and variable physical process.

The result was that many applications weren’t suitable for neural networks and others required a lot of prep-work. It might be necessary to rework your data collection system to get more and better data. It also was useful to pre-analyze the inputs to eliminate any dependent variables. Now, technically, that’s part of what the neural network should be good at – extracting the core dependencies from a complex system. However, the amount of effort – in data collected and training time – increases exponentially when you add inputs and hidden nodes, so simplifying a problem was well worth the effort. While it might seem like you can always just collect more data, remember that the data needed to be representative of the domain space.  For example, if the condition that results in your process wandering off spec only occurs once every three or four months, then doubling your complexity might mean (depending on your luck) increasing the data collection from a month or two to over a year.

Hopefully you’ll excuse my trip down a neural net memory lane, but I wanted to set your expectations of neural network technology where mine were, because the state of the art is very different than what it was. We’ve probably all seen some of the results with image recognition that seems to be one of the hottest topics in neural networks these days.

So back to when I read the article. My first thought was to think in terms of the neural network technology as I was familiar with it.

My starting point to design my own chess neural net has to be representations of the board layout. If you know chess, you probably have a pretty good idea how to describe a chess board. You can describe each piece using a pretty concise terminology . In this case, I figure it is irrelevant where a piece has been. So whether it started as a king’s knight’s pawn or a queen’s rook’s pawn, that doesn’t effect its performance. So you have 6 possible piece descriptors which need to be placed into the 64 squares that they could possibly reside upon. So, for example, imagine that I’m going to assign an integer to the pieces, and then use positive for white and negative for black:

Pawn Knight Bishop Rook King Queen
1 2 3 4 5 6

My board might look something like this 4,2,3,6,5,3,2,4,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0…-3,-2,-4

If I am still living in the 90s, I’m immediately going to be worried about the amount of data, and might wonder if I can compress the representation of my board based on assumptions about the starting positions. I’ve got all those zeros in the center of my matrix, and as the game progresses, I’m going to be getting fewer data and more zeros. Sixty-four inputs seems like a lot (double that to get current position and post-move position), and I might hope to winnow that down some manageable figure with the kind of efforts that I talked about above.

If I hadn’t realized my problem already, I’d start to figure it out now. Neural networks like inputs to be proportional. Obviously, binary inputs are good – something either affects the prediction or doesn’t. But for variable inputs, the variation must make sense in terms of the problem you are solving. Using the power output of a pump as an input to a neural network makes sense. Using the model number of that pump, as an integer, wouldn’t make sense unless there is happenstancial relationship between the serial number and some function meaningful to your process. Going back to my board description above, I could theoretically describe the “power” of my piece with a number between 1 and 10 (as an example), but any errors in my ability to accurately rank my pieces contribute to prediction errors. So is a Queen worth six times a pawn or nine? Get that wrong, and my neural net training has an inaccuracy built in right up front. And, by the way, that means “worth” to the neural net, not to me or other human players.

A much better way to represent a chess game to a mathematical “intelligence” is to describe the pieces. So, for example, each piece could be described with two inputs, describing its deviation from that piece’s starting position in the X and Y axes, with perhaps a third node to indicate whether the piece is on the board or captured. My starting board then becomes, by definition, 96 zeros, with numbers being populated (and generally growing) as the pieces move. It’s not terribly bigger (although rather horrifyingly so to my 90s self) than the representation by board, and I could easily get them on par by saying, for example, that pieces captured are moved elsewhere on the board, but well out of the 8X8 grid. Organizing by the pieces, though, is both non-intuitive for we human chess players and, in general, would seem less efficient in generalizing to other games. For example, if I’m modelling a card game (as I talked about in my previous post), describing every card, and each of their possible positions; that  is a much bigger data set than just describing what is in each hand and on the table. But, again, it should be clear that the description of the board going to be considerably less meaningful as a mathematical entity than the description creating working from the game piece.

At this point, it is worth remembering again that this is no longer 1992. I briefly mentioned the advances in computing, both in power and in structure (the GPU architecture as superior for solving matrix math). That, in turn, has advanced the state of the art in neural network design and training. The combination goes a long way in explaining why image recognition is once again eyed as a problem for neural networks to address.

Consider the typical image. It is a huge number of pixels of (usually) highly-compressible data. But compressing the data will, as described above, befuddle the neural network. On the other hand, those huge, sparse matrices need representative training data to evenly cover the huge number of inputs, with that need increasing geometrically. It can quickly become, simply, too much of a problem to solve in a timely manner no matter what kind of computing power you’ve got to throw at it. But with that power, you can do new and interesting things. A solution for image recognition is to use “convolutional” networks.

Not to try to be too technically correct, I’ll try to capture the essence of this technique. The idea is that the input space can be broken up into sub-spaces (in an image, small fractions of the image), that then feed a significantly smaller neural network. Then, one might assume that those small networks are all the same or similar to each other. For an image recognition, we might train 100s or even 1000s of networks operating on 1% of the image (in overlapping segments), creating a (relatively) small output based on the large number of pixels. Then those outputs feed a whole-image network. It is still a massive computational problem, but immensely smaller than the problem of training a network processing the entire image as the input.

Does that make a chess problem solvable? It should help, especially if you have multiple convolutional layers. So there might be a neural network that describes each piece (6 inputs for old/new position (2D) plus on/off board) and reduces it to maybe 3 outputs. A second could map similar pieces.. where are the bishops? where are the pawns? Another sub-network, repeated twice, could try just looking at one player at a time. It is still a huge problem, but I can see that it’s something that is becoming solvable given some time an effort.

Of course, this is Alphabet, Inc we are talking about. They’ve got endless supplies of (computing) time and (employee) effort, so if it is starting to look doable to a mere human like me, it is certainly doable for them.

At this point, I went back to the research paper wherein I discovered that some of my intuition was right, although I didn’t fully appreciate that last point. Just as a simple example, the input layer for the DeepMind system is to represent each piece as a board showing the position of the piece. So 32X a 64-by-64 positional grid. They also use a history of turns, not just current and next turn. It is orders-of-magnitude more data than I anticipated, but in extremely sparse data sets. In fact, it looks very much like image processing, but with much more ordered images (to a computer mind, at least). The paper states they are using Tensor Processing Units, a Google concoction meant to use hardware having similar advantages to the GPU and it’s matrix-math specialization, but further optimized specifically to solve this kind of neural network training problem.

So lets finally go back to the claim that got all those singularity-is-nigh dreams dancing in the heads of internet commentators. The DeepMind team were able to train in a matter of (really) twenty-four hours a superhuman level chess player with no a priori chess knowledge. Further, the paper states that the training set consists of 800 randomly-generate games (constrained only to be made up of legal moves), which seems like an incredibly small data set. Even realizing how big those representations are (with their sparse descriptions of the piece locations as well as per-piece historical information), it all sounds awfully impressive. Of course, that is 800 games per iteration. If I’m reading right, that might be 700k iterations in over 9 hours using hardware nearly inconceivable to we mortals.

And that’s just the end result of a research project that took how long? To get to that point where they could hit the “run” button took certainly months, and probably years.

First you’ve got to come up with the data format, and the ability to generate games in that format. Surprisingly, the paper says that the exact representation wasn’t a significant factor. I suppose that it an advantage of its sparseness. Next, you’ve got to architect that neural net. How many convolutions over what subsets? How many layers? How many nodes? That’s a huge research project, and one that is going to need huge amounts of data – not the 800 randomly generated games you used at the end of it all.

The end result of all this – after a process involving a huge number of PhD hours and petaFLOPS of computational power – you’ve created a brain that can do one thing; learn about chess games. Yes, it is a brain without any knowledge in it – a tabula rasa – but it is a brain that is absolutely useless if provided knowledge about anything other than playing chess.

It’s still a fabulous achievement, no doubt. It is also research that is going to be useful to any number of AI learning projects going forward. But what it isn’t is any kind of demonstration that computers can out-perform people (or even mice, for that matter) in generic learning applications. It isn’t a demonstration that neural nets are being advanced into the area of general learning. This is not an Artificial Intelligence that that could be, essentially, self-teaching and therefore life-like in terms of its capabilities.

And, just to let the press know, it isn’t the end of the world.

Known Bugs – Arab Israeli Wars

Arab Israeli Wars

Version 0.1.3.0. (April 22nd 2017)

  1. If you move a vehicle, and still have additional transfers left, but don’t want to use them, the system may wait forever for you to make another transfer. This doesn’t always happen, but if it does, moving a unit around within the same front will usually result in the prompt.
  2. If multiple informational displays are show simultaneously and overlapped, the hide button may show through the upper card from the lower card.