Game AI Pro
gameaipro.com
gameaipro.com
For AI researcher Game AI is like porn. It's cheap tricks and obviously fake but oddly fascinating. Sometimes you find a new trick you want to try in real life.
Marvin Minsky once said "I bet the human brain is a kludge." If I had to bet, I would say that human brain is full of dirty tricks, incomplete solutions, shortcuts and artificially limited problem spaces evolved to pick berries and avoid tigers, not to understand the world. Combining many tricks together can create illusion of generality that is very convincing.
Indeed, Game AI puts the "Game" first -- as it should for Games, but to the detriment of anyone who cares more about AI and wants Games to be a fun place to study AI.
If we split Game AI into "problem solving" (like pathfinding) and "opponent personality", then we can recover a lot of good AI that generalizes beyond games, without being misled by the parts that only useful for tricking people in a toy environment.
Wikipedia has an interesting list of cognitive biases [1]. Going through these, I tend to think of all of them as heuristic failures, where those shortcuts and incomplete solutions are pushed to edge cases.
However, I’m not a neuroscientist, and my knowledge and use of A.I. is limited to hobby projects.
It's perfectly possible our heuristics are muddied beyond necessity in order to generate variety in action, so as to reduce risk to the unknown for the species - even though it would cause a minority of individuals to consistently make sub-optimal choices. From a speculative standpoint, it's easy to find examples of people doing things that we consider 'stupid' but it pays off because of some unlikely event occurring in coincidence.
Reality has a lot of unknowns. There is no perfect model that could account for that. It's possible being hyper-intelligent (beyond our current ability) is (or was) a disadvantage for the species
Neuroscientists are the new doctors of the 50s. We thought the appendix was useless turns out it has many uses. We thought priming and all these “tricks” were things and then the crisis and Kahneman et al were debunked.
I bet the brain is just as elegant and powerful as it has to be to do the incredible complex things we do, and we’re just so far from really understanding it that we run around appendicizing all sorts of things we just don’t really know well yet.
But no more than that. Which is what Minsky was saying.
Sorry but you’re not following my point or his if you think they agree.
Why can't it be both? Maybe the duct tape and shortcuts are what give the human brain it's fantastic adaptability. Duct tape and shortcuts aren't a bad thing necessarily. Personally, I think that's what gives the human race as a whole it's fantastic adaptability. You've got millions of people each with their own duct tape and shortcuts to the same problems, meaning each of us does things just a little bit differently, we see other humans with their shortcuts and slap them onto our own with some duct tape and we get better at things or learn something new. Do this over millions of years and generations and you've got a pretty damn capable brain that's slapped together millions of years worth duct taped together solutions and skills that keep growing as we hand our giant ball of duct tape to successive generations.
It is funny how willing computer scientists are to want to use a duct tape analogy. I think it’s because programming, which is basically the polar opposite of a brain (precise and unintelligent) requires so much damn duct tape if you want to get anywhere useful. Meanwhile a brain literally requires as little duct tape as possible if you want it to be generally good.
I'm a hobbyist at sleight-of-hand magic and therefore like to read and watch lots of material on the matter (eg the book "Sleights of Mind" by neuroscientists Stephen Macknik and Susana Martinez-Conde, and science writer Sandra Blakeslee) and there are an incredibly many ways to fool or trick the brain into believing untruths, even obvious ones.
Cognitive biases were also incredibly exaggerated and if you’ve read into the replication crisis and various pushback Kahneman has gotten on his studies you’ll find much of what we call cognitive bias was wrong or exaggerated.
That we figure out super inventive ways to trick others into thinking we’re not smart is just more proof of how clever we really can be.
A great example is loss aversion which, if you read the debate around, is a totally smart strategy given we have absorption barriers.
That I pay attention to things in a certain way that allows magicians to fool me in predictable ways doesn’t mean our brains are failing in some way - it means they are making a trade off that helped us survive better with the given processing power we have. But that trade off is almost certainly the smart trade off, and labeling it “kludgy” is lazy thinking.
You missed the point of what I was saying, which is that because of my interest in sleight of hand, I spend a lot of time reading and watching different material on "magic", the book Sleights of Mind being such material. Its not about sleight of hand, but rather that it's easy to trick the mind into believing things. Also, most sleight of hand isn't really about what our eyes see at all, but more about masking actions -- hiding one action with a bigger action or through misdirection, making use of the fact that our minds find it hard to focus on multiple things at once. I would say most sleight of hand tricks are about curating what the other person perceives. Sure, sometimes its by hiding what is being done from the eyes, but often its doing something in plain sight but in a way that the spectators don't connect the dots.
From the wikipedia synopsis of Sleights of Mind: "Macknik and Martinez-Conde say that magic tricks fool us because humans have hardwired processes of attention and awareness that are hackable. Good magicians use our inherent mental and neural limitations against us by leading us to perceive and feel what we are neurologically inclined to."
I'll give you a simple example:
In card magic, using a concept known to magicians as "time misdirection" (that is, by putting time between cause and effect), spectators often retell the performance as having happened in their hands, even when they never actually touched the cards.
The book and other material has more. The point is that our brains piece together incomplete information and make assumptions. We see what we expect to see, hear what we expect to hear. Our brains pattern match and see patterns in noise.
I'm not saying that this isn't useful to our survival, if you look at the outer ring of this image from wikipedia[1], it shows that there are very good and valuable reasons why we do these things, but they still lead us to make mistakes, errors, wrong assumptions or act illogically, sometimes to disastrous effect. I'm not and didn't say it was a "kludge" exactly, but I do think that the brain is taking less-than-perfect shortcuts (to save time, memory or make up for too much/too little/too noisy information).
[1] https://upload.wikimedia.org/wikipedia/commons/6/65/Cognitiv...
I'm not a computer scientist. I'm just going off of the way I learn and the way i've watched other people learn or the way i've seen people learn while i've been teaching them or training them. That and the general way everything kind of works. As amazing as everything in the world seems, when you break everything down to the smallest components, they all rely on the same skills and techniques we've been using since we were picking berries and hunting mammoths. The materials, accuracy and scales of our work have changed and improved dramatically, but fundamentally, most of what we do can be traced back to the same old things we've always done. We're just really good at building on layers and layers of things and applying things and knowledge to novel concepts and ideas.
Take music for an example, we just keep making the same music over and over and over again, yet we still find new and novel ways to make it sound different and new to the point where most people don't realize they've been listening to the same few songs in new forms for the last 100 years or so at least.
What do you mean by this? Oftentimes I see people reaching the conclusion that because a lot of modern music is built around the same major scales and largely homogenous chord progressions, it must be the same, but this simply isn't the case.
But it is, to the point where there's songs I learn only somewhat and i'll get confused to which lyrics are which and I'll sing a mix of the two songs when playing them, imagine and no woman no cry jump are two it happens to me with all the time. Same with truly madly deeply and kryptonite oddly enough. Then there's the whole pachabel's canon meme which is entirely true. There's a few songs, the most recent is some maroon 5 and i think a jonas brother's song which I was confused as hell as to why I liked until I realized they're just another rework of canon.
That's not even getting started on the direct ripoffs such as the 1000'@ of songs based on the Amen break or the thousands of songs that are basically a simple 8 or 12 bar blues progression or the tangled web of constant remakes that is reggae and dancehall and every song that lifts a bassline or melody from them without credit all the way up to such obscure things as the friendly neighbours tune from earthbound being a relick of the real rock riddim.
It’s a heuristic that can be wrong, but the heuristic itself isn’t kludgy just the occasional specific use case can be wrong.
Here’s an analogy: any individual ant does many things that look stupid for its own survival. But would you call the ants thinking “kludgy”?
Can you elaborate?
I recently learned that current appendectomy procedure (at least in Europe) calls for the removal of the appendix even if a surgery should reveal that the organ is in fact completely healthy – because the procedure from that point is considered so unrisky, and the organ considered so useless, that the potential future risk of a medical professional misinterpreting the existence of an operational scar as an indicator of a previously performed appendectomy is deemed more problematic.
Can we still cite Marvin Minsky in the AI field given the allegations that have arisen regarding his relationship with Epstein and sex trafficking?
We should find better luminaries.
I'm all for redemption and I hope Stallman comes around and apologizes, but Minsky went to the grave with whatever happened.
Yeah, that's the thing - the brain and body evolved for a much different environment than the one we find ourselves in today.
So some aspects which used to be advantages are now disadvantages that we have to actively manage with things like exercise, healthy diet, meditation. Which the brain fights against.
And even though they gave a talk about in in 2015, their "Simplest AI Trick in the Book" is still not implemented by some games released nowadays.
In case you don't know it, it's:
0.2s reaction time for aiming
+ 0.4s reaction time for yes/no decisions
+ additional delay for ambiguity, surprise, or limited visibility
I wholeheartedly agree with this advice. Just seeing your opponent taking a moment to think makes whatever it is they do so much more convincing.
wget -r -A.pdf http://www.gameaipro.com/
wget -r -A.zip http://www.gameaipro.com/
a) sell PDF ebooks
and
b) explicitly say that you are not supposed to re-upload the content elsewhere.
And load does not translate to cost for everybody. If you saturate the connection to my VPS, I don't pay more, it just gets slower for everybody in contention. I can spin up mirrors but if I'm offering a free resource like this, I'd be more likely to limit the bandwidth-per-client-IP or just actively let it run slow. They could even limit the bandwith to the subdirectory with...
location /download/ {
limit_conn addr 1;
limit_rate 50k;
}--wait=5000 --random-wait
Done...
Very cool how recent and modern these are (along with super reputable authors)
I just finished the first four sections and I love it. Thanks a lot!
I think it might give the gaming industry a kick in the pants to start utilizing more advance AI techniques in general, since it seems almost all discussions of strong AI in games are dominated by apologists explaining why it's not practical. Just one example of strong AI in a successful game would change the industry.
After strong AIs are common, we can persue the even more interesting task of dumbing them down in fun ways.
An artificially handicapped or limited smart ai in video games is often obvious and not fun.
Starting with an AI that is weaker than human players, and then artificially limiting it even further does not produce a great AI. Although, some players still choose to play against these easiest of AIs.
I believe starting with an AI that is far stronger than any human, and then artificially limiting it will be fun. This is a fundamentally different situation. This is like a chess engine, no human can hope to win without some artificial limitations on the AI. This is a type of AI we have never experience outside of a few abstract board games like chess. I look forward to seeing AIs like this come to 4X style games such as Civilization.
> Game engines have typically made reproducing these types of problems easier using deterministic playback methods (Dickinson 2001), where the entire state of the game simu-lation can jump back in time and resimulate the same problem over and over (Llopis 2008).
Imagine if you could do this for all programming? [from chapter 6]
rr comes pretty close. https://rr-project.org/