BletchleyGeek

Thanks @BFCElvis for the update, the Italy games have always had a temperamental TO&E, so I keep my fingers crossed that the tidying up doesn't turn into a death march.

I was about to use words to that effect. He's a very cautious and meticulous player. We played a game of the Market Garden Devil's Hill scenario a few years back and it took a while to find stumble upon his troops.

That paragraph could totally be in a book by Julian Barnes Nicely written, Ben.

"Listen up!" ?

Looking forward to see and hear the Indian Army in its full digital glory... they fought probably the hardest and toughest of campaigns in Italy and Burma. @Warts 'n' all I am personally looking forward to an Scottish voice pack.

Interesting stuff, thanks @BFCElvis.

Not on the commercial engines, but Graviteam's engine does.

I was about to ask "where is it set" but I will wait for the DAR/preview (if that's happening this time around).

For creating colourful, diverse animations, yeah, that is for sure. One of the issues in applying AlphaZero to professional wargaming is dealing with the problem of time, space and coordination of assets to, say, evaluate possible COA implementing a Strike mission. Random exploration, which is what AlphaZero does, has exponentially diminishing probabilities of landing on a particular game state via a valid set of moves. In a game with two choices, the probability of a random exploration heuristuc generating a particular sequence is very easy to calculate and vanishes to zero as the length of the sequence grows. It worked for Alpha because games have nice terminal states where you get a very simple payoff, and you can extract easily features that correlate well with possible game outcomes. David Silver had a lot of work with handcrafted features before turning to CNNs. And still, having to play games to completion is probably the most time consuming part of the execution of the Alpha Algorithms In other settings you need a heuristic, or base policy, to show the way. That is, old good "AI" techniques whose practitioners call model based search and optimization...

I have been ignoring Rattenkrieg for a while when became apparent he was not interested in having a technical discussion. There is, in my mind, a very clear difference between promising and delivering, and between mocking and challenging. I sketched a very reasonable benchmark to test Rattenkrieg's assertions, and asked for a budget. Research does not happen for free, and what Rattenkrieg proposes is a research project, not a software development project. The experience of the Leela project is well documented, and the challenges of integrating modern machine learning. Maybe he thinks I am mocking him because he has an idea of the costs involved. There is active research on applying deep learning to professional wargaming. The problems that are being found are lack of availability of computational resources to match results within reasonable time frames, issues with the asymmetry of military wargaming scenarios that do not fit well with the well behaved and neat game theoretical structures of classic boardgames, issues with representation as CNNs cannot exploit high dimensional state representationsto come up with generalised features, and more, like the problem of composability I discussed earlier. People have been trying quite hard on this for a few years now. Contesting wild claims by a company is not "over the top". We should all do it more often. But I have no "friends" in that company either. The USARL has been working for a long time on mapping out the challenges posed by the so called the Internet of Battle Things https://arxiv.org/abs/1712.08980 So let's say that, considering the claims made by Rattenkrieg, and looking at the challenges, I remain skeptical that either Palantir can solve that problem, or that they need to solve that problem in order to provide the US Army with next gen information processing and communications systems. Confronting the CV community with proof that well known approaches to object recognition suffer from massive overfitting was probably not going to go down well, but the facts are that they have been shown to break down when inputs are trivially perturbed with noise. Hence why they are not suitable technology, by themselves, to provide automatic target acquisition and I stand by my remarks. They can be easily spoofed and can pretty much track anything - see the examples on the papers I linked on the first post. So much for "sweeping statements" really. There is even a new field of research, very hyped up, called "Adversarial Machine Learning" just studying this, with varying levels of success. The locomotion suite by Google is a very useful research platform. Still, there are fundamental limitations on what kind of tasks can be learnt by neural networks without relying on specialised architectures and carefuk selection of parameters, search and optimization algorithms to guide and implement the training process. The work on Universal Planning Networks illustrates this. The locomotion results are relevant for the video games industry, and probably there is already work published on SIGGRAPH about it. Having an animated 3d character go from adjacent nodes in a navigation mesh is a remarkably easier problem than having humanoid robots going up the stairs of a randomly chosen house in America. The videogame developer has literally godly powers to shape physics in such a way that everything runs in real time and looks goid enough. And with this I have pretty much said my piece. I hope some of the folks here found this readable and informative. If anybody wants to know more about any of the above they can PM me or ask for further discussion.

Always a good cartoon to keep in mind @sburke. Thanks for that post, too

There is a big difference between gaining a contract and delivering a product that works as expected. It also does not seem to be claiming to provide what Mr. Rattenkrieg says it does, or definitely not one of the main deliverables. Somebody just pulled out the technical card on another forum member, and made arguments and claims which seemed to me to be both misinformed and hyperbolic. I pointed to this person how those claims were at odds with the facts. Here I am wearing my scientist hat, not my BletchleyGeek the forum member, as I think it is my civic duty to contest claims that I perceive to misinform the public. As usual on the Internet, this person just run to the top of the hill ready to die on it while refusing to engage, not actually answering any arguments or providing any reference. That is a waste of time and a dissapointment, as the discussion went onto technical details. I respected him enough to present some substantial arguments to discuss, and being way more intellectually generous than him. Thanks very much for calling me a 12 year old and a nerd... how is that supposed to be helpful? I do not suffer fools gladly when talking shop, and if you couldn't follow the discussion I am sorry you felt left out and excluded, but there are limits to the amount of time one can put into being didactical.

Cheers @IanL, happy to hear someone found the discussion interesting. The technology is there, and in a very useable state - much more than it ever was in the late 1980s and 1990s when first wave of neural networks practical algorithms and applications came up. Here's another paper you may appreciate reading (and sorry if I am making too many assumptions about your background: https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf

For the benefit of other readers interested on what it takes to reproduce the success of the Alpha algorithms it is useful to consider that Deepmind reported to beat Stockfish, a state-of-the-art automated chess player which has been developed for ten years, in just four hours (link to paper). That's impressive but it is also useful to remember that during those 4 hours, AlphaZero was trained quoting from the paper. A distributed, open-source effort to replicate those results against Stockfish took over a year, painstakingly following Deepmind approach https://en.chessbase.com/post/leela-chess-zero-alphazero-for-the-pc and they released it as a plugin for Fritz. Having a beefy GPU for computing the opponent moves is highly recommended.

This sentence proves that you just didn't understand anything of what I wrote initially, or that you know much or anything about how to apply deep learning to any practical problem from scratch, or you have read any of the papers. AlphaZero was initially demonstrated on the game of Go, and then Deepmind renamed the approach as Alpha or AlphaX where X is the name of the game they have deployed the same algorithm and changing the model of the game rules and states. The "zero" comes from not using games with humans for bootstrapping, which have zero to do with all the other questions I raised, which are all design choices made by humans. Don't bother coming back. Save your time for productive endeavours.

Okay, you actually helped me to parse a sentence of the manual that has beaten me for years - as anybody can see there's no mention of having to press the LMB at the same time as the Shift or Alt key.

I was talking about AlphaZero (or the Alpha style game players as they have gone on to play many other boardgames, including chess). The same applies to the more recent work on multiplayer first person shooters. They are well aware of using domain knowledge encoded as I said, I had no problems discussing that in person with Deepmind researchers in conferences, or watching them live answering to much more pointed questions from the audience in said conferences. Unbounded levels of aggro and ****ty attitudes are more often found on the Internet than in real life. Please educate yourself a bit for instance studying recent work like that on Universal Planning  Networks and how those networks are constructed for another example of what I was saying. And still, nothing of what you say addresses my comment pointing to the search of parameters, connections patterns, computation power requires, access to a model, generalization in neural networks, implications when deploying on the workflow of a game development company etc. Not to mention that one of the main components of AlphaZero/X is Monte Carlo Tree Search, an algorithm which was invented in 2006, and not learnt from interacting with a simulator. So zero actual answers, zero references to papers or actual demostrations, just more e-peen swinging and accusations of being butt hurt about something (actually, some of the general algorithms I have worked on for the ATARI games haven't been beaten yet by any of the neural stuff). Your answer is disappointing to me, as it demonstrates that trying to engage with you on technical terms, while presenting an intellectual challenge at the same time, as there is no debate without some friction, is just a waste of time. If you wanted to prove me wrong, you just need to produce that system that learns to do a squad level drill from just capturing the sequence of frames rendered and provides inputs via, say, the same way people cheat on MMOs with bots and @Japanzer used to automate map creation years ago. I would love to be proved wrong, as that means that I would learn something from it. Just go an budget it for us, Battlefront will be very happy to learn that is doable with limited resources without running at a loss for years as any research project usually does (and almost always stay there). Thanks very much for the links to the Graviteam manual, but my question was about splitting squads etc being available on all game modes. I can't seem to get it work when I am playing with the "battlegroups" variants of the campaign.

On the topic of clunkiness: it is clunky that I have been played Graviteam's games since 2011 - not sure I have clocked 500 hours, as I have to work some time - and I have precisely f*ck all idea about being able to do this after reading their "manual" and doing their "tutorials" several times over the years. Or is it just possible to do this on the quick battles? Plenty of clunky arbitrary design decisions like that in GT... as in pretty much every game out there as there no such a thing as a perfect design.

Just some notes on the above, as a member of the research community, I feel I need to comment on this briefly. First of all, that stance of "anything rulebased is anything but AI" is a disingenous position that does not hold when contrasted with the state-of-the-art literature. Domain knowledge can be expressed in many ways: with if-then statements, or behaviour trees, all the way to neural network architectures engineered to "capture" very specific processes and signals. The successes that have been widely hyped up - like the Star Craft/DOTA players by Deepmind and OpenAI - are fundamentally hybrids of what you refer with scorn as "AI" and machine learning. Even the Alpha players rely on not an insignificant amount of handcrafted knowledge, from the basic features used to parametrize states to the particular selection of activation units and interconnection patterns. All those choices were made by humans seeking the best combination of parameters, architectures, initialization strategies and more. If you check the paper on AlphaGo on Nature you'll see that the section devoted to explain those details is actually longer than the main paper. Even more interesting is to see how former preachers of the "end-to-end learning" gospel are now turning to classics like early 1980s subsumption-like architectures to bootstrap and guide those neural networks training process. Suffices to say that all major companies working on self-driving vehicles have abandoned that gospel and are scrambling to snatch leading researchers of areas which two years ago were considered to be "not relevant any more". I have no idea what is Palantir trying to pitch but it sounds to me as pure bull**** tbh. This is for several reasons: tactics require to deal with partial information, considering processes that flow at different time scales, on environments which are complex and dealing with a wide variety of platforms that operate autonomously (e.g. single riflemen, AFVs, a drone and its controller). At the contrary than in games like Go, where the number of pieces is fixed and known, and the board is always the same, a contemporary, near-future or past tactical environment shares little or none of those features. The most fundamental issue - there are several, and there's plenty of fundamental issues to choose and work on - to me is that neural models are not composable. That is, you can work out a neural network to say, steer a simulated squad of simulated robots broken into two teams just all right along a given line of advance and against a specific amount and direction of enemy fires. Here is a list of the dimensions such neural network has to generalize in order to be useful and interesting: - Initial distance to target (assuming that the order is to Assault) - Effective volume of fires on enemy positions as distance to them changes. - Type of terrain the unit maneuvers. - Obstacles obscuring LOS and LOF - Equipment of unit - Hypothetical equipment of the enemy There is absolutely zero guarantee that a given NN that performs at a certain level, for any meaningful performance index, on a finite sample along these directions will generalize to any possible combination of the above. If you have an algorithm for that which you can use on any problem at hand, then congratulations, you probably solved too Hilbert's 10th problem. This applies to everything, including Starcraft: how many possible Starcraft maps there are? Can you classify all possible tactical and strategic situations neatly into discrete homogenous categories? That's also why doing funny stuff to allegedly state of the art CV pre trained networks - like adding a 1-pixel wide black border to an image - catastrophically degrades the accuracy of object identification. Luckily, other than perhaps Russia and China I think, nobody even considers to deploy deep learning systems for target identification and acquisition. If somebody does, they're criminally insane or selling snake oil, or both. Provably you haven't done any of the above, but you may have a quite decent closed loop control strategy that works well enough to make some nice videos to impress people, or even beats some hand coded controller that somebody put a decent amount of effort in designing exploiting knowledge about the laws of Physics or some other fundamental process. That can be good enough, it all depends what you're comparing it against. Definitely you can't make any guarantees on suitability for any purpose other than that captured by your training set: YMMV. The problem of composability is illustrated by the following question: can I use that neural network as a building block to coordinate the movements of a platoon? The answer, so far, has been a quite deafening no. There is no known way to constrain back propagation to guarantee that the knowledge acquired by the neural network you are using as a building block is going to be obliterated or changed in a fundamental and undesirable fashion during training for the "composite" problem. Composability also challenges the ability to train incrementally, as the capabilities of the unit change due to casualties or changes in equipment. There's again no guarantee that any knowledge will be preserved when re-training after changing those elements in the environment that generates the training data for the neural network. Last, composability has to do with time: what is the minimum period of time to be considered? Is there a sensible upper bound on the number of such consecutive periods of time to consider? Taking off-the-shelf techniques used for Natural Language Processing has been shown to be pretty much like dancing about architecture, spoken and written word has a very definite temporal structure, for which we know its "laws" (because we invented grammar and rules of style!). Another fundamental problem linked to this last observation is that whatever the neural network learns we cannot be sure that it is capturing the essential first principles that allow the behaviours which are to be mimicked. This is analogous to the fundamental issue with the classic research by T. N. Dupuy and the HERO institute - in the 1960s, one could overfit a model only by hand, in the 2020s you can use neural networks too! Contemporary machine learning has a niche, like those "rule-based" approaches you disparaged in your post do. And I certainly appreciate the good things in deep learning, for instance, the dependability and efficiency, provided that the right conditions for the techniques involved to work properly are an invariant of the set of situations I need to deploy them. Going back to the games briefly. Regarding Graviteam, I learnt through a weird interaction with Andrey on the Steam forums a few months ago that he's pretty ignorant on any of these topics. Which is totally all right, he's not expected to be an expert on that. So my educated guess is that what you see animating those pixel truppen in Graviteam games are not too different from the techniques used in 99% of video games and 80% (?) of robotics: good old hand-designed controllers via behaviour trees, A*/D* and PIDs/SQP/Non Linear Programming. Last, I want to address the comment which I read is blasting BFC (and video game developers in general) because of not using deep learning technologies. I have zero idea of what is the operating budget of BFC, but say, the cost in $$$ to say develop and train an Alpha-like system for one of the countless drills possible in CMx2 would probably be somewhere north of 1 million USDs (that counts salaries, on boarding of staff and compute for like 40 days with a similar amount of computation power as the one wielded by Deepmind to ensure you can beat Bil Hardenberger like 90% of the time). Indeed, they would probably amortize salaries and onboarding over time, but the cost of computation is what it is, and changes in the game mechanics, or even bug fixes, etc. would require retraining (or training new networks for that special case). Indeed there are opportunities for more modest applications rather than end-to-end tactical battle management, but I am skeptical than they are cost effective for the return on investment Battlefront will get. I am pretty sure they're already doing this pretty much for the sake of the arts, and unless they get patronage, I can't see why should they spend tens of thousands of dollars per month on EC2 just to replace their code for animations, drills etc by neural networks. Or maybe you could work pro bono for Battlefront developing those

We are pretty much in agreement re: first paragraph. On the second paragraph I am more skeptical because I can see that working when surprise is achieved, or again infantry without heavy AT weapons caught in the open. Then armour pretty much has the effect of cavalry before the introduction of automatic weapons. The 45mm gun was in principle capable of defeating the armor of most early war German tanks if not all, and AT rifles would be dangerous against sides etc. Of course, there are many cases when heavy weapons were scarce/not available, or troops weren't very inexperienced, or tactical leaders were quite incompetent. But I don't think thst combination happened "any given Sunday".

Effective AFV employment requires to concentrate them. You need mass your AFVs but you do not need to attack the enemy frontally. The Red Army had plenty of anti tank guns, but few commanders that knew what to do with them - the experience of the Spanish Civil War notwithstanding. That's why there are documented examples of an Army commander instructing the divisional commanders where to place their batteries. What changed in 1943 is that besides having better guns, for the first time you got to see German armour channelled into defenses in depth with carefully thought out kill zones. Which was pretty much the unsolvable tactical problem the Soviet armoured forces had to deal with for a really long time. On more fluid situations it all boils down to good command, control, training and luck. Andrey, for whatever the reason, thinks that the frontal attack on broad fronts the only tactic used ever by the Red or the German Army. I still enjoy playing the AI but I have to restrain myself from setting up traps as the AI will press on frontal attacks while being attacked on the flanks. Last week I annihilated one German infantry battalion just like that. Maybe it is a problem of how tactical battles are setup, but I have rarely seen the AI pulling out a concentric attack. Of course, it is always simpler to take Andrey's approach to problem solving: say it does not exist and then eventually fix it a few years down the road...

Very true, and a pretty good line too, John.

Thanks @DMS. It is true that armored tactics weren't that sophisticated sometimes, especially for badly led and trained formations. But not all the fighting was like slapping someone with a leg of ham in the face

Good summary @Thewood1 - but here I would say that they put their mind to it, but their mind goes to places ours does not. I can't read their Russian-speaking forums, I would love to hear if the feeling is similar across the language divide. Maybe @DMS can chime in?