Rattenkrieg

For the curious on how AI is learning to play video games:

Is it really 2 guys developing all of this? Vehicles, animations, UI, textures, sounds, MP infra etc?

Thank you! @37mm Indeed I was able to get the translucent foliage issue to disappear by using the Advanced War Movie shaders, which seem to disable MXAO. One thing that is a bit off-putting is the degree to which metal reflects light to the point where the entire front of an AFV can turn white at certain angles.

Thanks @37mm that makes sense. Could you post your shader preset that you use for the WW2 games?

Any new Reshade presets or tips? I noticed that turning MXAO on makes low walls basically burn through all foliage, creating a series of criss-crossed lines on the screen. Does anyone else see that?

Hello folks - wondering if anyone knows how to increase the draw distance outside of the max settings in the game? Sometimes in other games I have been able to set up config files to override game menu settings, with things such as forcing memory sizes, screen resolutions and draw distances. Thanks!

Thanks guys, much appreciated.

I couldn't have come up with a better coda. Here are 2 fun little nuggets for the curious. 1) Build your own Terminator Target Acquisition HUD in Microsoft Hololens. 2) Where did that "code" on the original Terminator HUD come from? std::terminate

Ian, but I am discussing AI and DL with you. The Nvidia video and paper are incredible advances that should get game developers very excited. Imagine being able to use a GAN-based SDK from Nvidia to create a photo-realistic WW2 game based on DL from 10,000+ hours (x25fps) of combat footage from WW2 that has been restored by AI to 4K? Imagine being able to train your "pixeltruppen" using Google's (DeepMind) autonomous locomotion framework for rich environments (when it is released)? Just because two people who work in/with AI don't like how the other behaves on a forum doesn't mean there isn't already ample material here for enrichment.

I struggle to see the connection between the suggestion they use data and the inference that there is a recycling bin approach to it. I suggest they start gathering gameplay data and analysing it, because they generate an enormous amount of it (every bullet, footstep and click) and have the competitive advantage of H2H and registering what players look at when they replay WEGO over and over. That's valuable. BFC has been around for 24 years is it? I am sure they will want to be around for another 26 at the very least. They could reach out to universities and get interns who love military, simulation and AI, or a combination of those interns, or various military schools throughout the US (assuming they are a US Inc.) They could apply for an SBIR grant, like this startup did for computational video editing and was awarded $224,734: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1842850&HistoricalAwards=false Here is an example of a resource that my interns are using right now to restore old videos: https://xinntao.github.io/projects/EDVR I chose not to get bogged down in a ML debate about the exact specifics of how a wargame could begin (note: begin, not deploy) exploring the possibilities of DL and AI in general. Firstly, BletchleyGreek mocked me by suggesting I work pro bono for BFC. Secondly, he made a spurious claim about a company where friends of mine work. Neither of those actions merit any response from me. He is comfortable in his knowledge of ML and his cognitive biases and he went on the attack from the first paragraph of his initial response, at which point I decided I will not respond. I'm sorry if that disappoints you. But why wait? Research is the companion of development and DL is evolving so fast that it's better to start sooner rather than later. @IanL yet another sweeping claim. Again, it betrays an embryonic understanding of DeepLearning that drastically limits my willingness to engage. Perhaps he will say he meant "deploy them right now, today." But isn't that what I am being accused of insinuating with my encouragement of BFC to look into AI? That they are insane for not having already done it? On the subject of DL systems for target identification and acquisition: https://digital-commons.usnwc.edu/cgi/viewcontent.cgi?referer=https://www.google.com/&httpsredir=1&article=5125&context=nwc-review Although the entire paper is very interesting, Section 3 discusses autonomous technology in targeting in more detail.

p.s. just in case I am not posting enough Arxiv links to back up my computer generated street cred: https://arxiv.org/pdf/1808.06601.pdf

The TL;DR version. My original statement above hypothesizes a future in which "vast amounts" of gameplay data and computer vision data derived from actual combat footage are processed by NNs to construct autonomous tactical models. It's a forward looking hypothesis of a direction in which wargaming AI may evolve, and something game developers should be encouraged to explore, because I can guarantee you that a day will come in the not-too-distant future where CGI does not mean models built by people. Having seen a Palantir demo in person (which I am sure naysayers will call a closed-loop hype demo based on bull****) of battlefield AI, I consider the convergence of these two fields to be simultaneously inevitable and exciting. This is the kind of work I am involved in (we are a partner but I do not work for Nvidia). What you see in the video below is a world that is constructed by NNs (based on a Generative Network) using Unreal 4. Now just imagine a world where instead of people driving the car through the virtual world, it's an AV and the other cars on which the models were built are also AVs. What starts as pure research will have immensely impactful practical applications within the next 5-10 years in peacetime cities and on battlefields. My first encounter with Computer Vision in business was with a hedge fund that rented several thousand apartments around the world. Each apartment overlooked a street with specific retail brands on it. In the window of each apartment were a multitude of video cameras that recorded every human that entered into and exited from each store. There were many thousands of cameras providing massive amounts of data and a highly accurate predictive model of the retail performance of each brand was derived from the CV data which included the size, colour and number of shopping bags that each identifiable shopper (anonymously tracked) emerged with, contrasted with their profile as they entered. It was also possible to track repeat shoppers even when they changed their clothes, headgear, hairstyles, etc. The hedge fund using this technology was able to consistently beat market predictions for the stock performance. That was nearly 8 years ago. I have the privilege of working in a pretty cutting edge field and I encourage everyone out there to familiarize yourselves with what's coming.

You are an exercise in self-contradiction. I stated that the AI being developed by Palantir as part of their winning DSCG-A bid is "the closest thing I have seen so far" to AI and CV being able to generate autonomous battlefield C&C. You jumped on your high horse and started telling me how wood and trees does not equal forest. You actually state that you have no idea what they are doing but that it is pure bull****. Have you any idea how silly that is? Yet, by your logic, I'm misinformed and hyperbolic? This means you have no idea about - yet simultaneously understand and are up to date with- all of the below. Palantir is eagerly awaiting your CV. SPECIFICATIONS ... Tactical Intelligence Ground Station (TGS): Tactical geo-intelligence PED and targeting node. TGS retains Common Ground Station capability and functionality; upgrades the hardware; adds more moving target identification (MTI), full motion video and imagery exploitation capability; and provides totally integrated stand-alone imagery, MTI and video sensor processing Geospatial Intelligence Work Station: Provides geospatial and imagery analysts within tactical and operational Army units the ability to process, view, exploit, transmit and store geospatial and imagery information via Army area communications from brigade to echelons above corps Operational Intelligence Ground Station: Consolidates the capabilities of the AN/TYQ-224A, GUARDRAIL Ground Baseline and the Tactical Exploitation System Forward Intelligence Processing Center: V1 provides a suite of core PED applications for intelligence analysis and storage. V2 is the basic combat training and division commander’s primary ISR networking; analysis, production system for tasking of sensors PED support https://asc.army.mil/web/portfolio-item/iews-dcgs-a/ And the work of all of the following involved: Lockheed Martin (Denver, CO) General Dynamics (Scottsdale, AZ) ViaTech Systems, Inc. (Eatontown, NJ) Palantir (Palo Alto, CA) MITRE (Eatontown, NJ) Booz Allen Hamilton (Eatontown, NJ) Raytheon (Garland, TX; Arlington, VA) BAE Systems (Arlington, VA) NetApp (Sunnyvale, CA) VMware (Palo Alto, CA) Esri (Redlands, CA) Tucson Embedded Systems (Tucson, AZ) L3 Communication Systems (Tempe, AZ) Dell (Austin, TX) Potomac Fusion (Austin, TX) Redhat (Raleigh, NC) IBM (Armonk, NY) HP (Palo Alto, CA) Leidos (Reston, VA) ManTech (Fairfax, VA) Oracle (Redwood Shores, CA) Microsoft (Redmond, WA) I have seen a demonstration of the following: moving target identification (MTI), full motion video and imagery exploitation capability; and provides totally integrated stand-alone imagery, MTI and video sensor processing. As I stated in my OP, I work in the field of AI and Computer Vision. I fight Palantir on a weekly basis to stop them from headhunting my people. You work in the field of ML and are applying, in my opinion, myopic and narrow thinking in your knee-jerk reaction claim that no self-respecting AI specialist would make unless they were frustrated at being bypassed. The completely unsubstantiated discrediting of a company perfectly fits your definition of a claim that misinforms the public. If you are, indeed, a scientist then it's even more shameful. So that's what you call adding a disclaimer that explains why I object to the overuse of the term "AI"? Again, your logic is insane. You complain that the results of DL experiments such as AlphaZero are overhyped but you are fully OK with the overuse of the term AI to mean anything that is a rules-based decision making framework? Reading your posts is truly a stultifying experience.

Well @IanL I use it as a litmus test. If you claim that, arguably, the world's leading AI defense contractor is "full of pure bull****" don't you think that a detailed set of reasonings is required to back up that statement? He lists a set of reasons as to why ML is, in general, difficult to apply to the dynamic complexities of combat, however when you make an outlandish claim that the established leader in battlefield AI (just ask Raytheon who lost to them) has completely fooled the Pentagon, that is akin to stating that you know exactly what the Pentagon put out an RFP for, what Palantir demonstrated, and why it was accepted. And that the entire thing is a ruse to defraud the US taxpayer. That is an enormously arrogant and irresponsible claim, and perhaps it is lost on you and most other readers. Was I, in fact, "overstepping" when suggesting that the devs start to train NNs to test their capabilities by gathering data? You may say that if you don't understand the potential of DeepLearning. The problem seems to be that BletchleyGeek is an ML researcher (I doubt he is a DL researcher) with a cognitive bias - and the difference is not to be underestimated - and expects me to rebut his statements when in reality the burden of proof is on him to prove that Palantir (and by extension the US DoD and CENTCOM) are full of "pure bull****" in the context of AI for warfighting application. The irony is that he then goes on to say just how amazing DeepLearning actually is, however nobody seems to notice the evolution in his position.

It's all well and good to post links to papers (just FYI the majority of the most interesting papers published on AI in 2018-9 are in Mandarin). I'm still waiting for @BletchleyGeek to back up his claim that Palantir ($800 million contract to build a DeepLearning based warfighting system for the US Army) is full of "pure bull****". How anyone can make a ludicrous claim like that and get away with it speaks volumes.

I do find it remarkable how incurious a researcher you appear to be. How could this have "beaten" you for 8 years and yet I figured it out so quickly? The [LMB] is used to "select" in all cases. So presumably [Alt] and [Shift] are modifiers of the "select" action. I had the manual open and clicked on everything on every page. I must be a machine. You couldn't figure this out or research it on the forums.

Makes outlandish, arrogant and boastful claim with zero evidence to back it up. Pops smoke. Retreats.

It's amusing to watch you contradict yourself. You mentioned Alpha then immediately how AlphaGo was trained by humans and absolutely no mention of AlphaGo Zero. How am I to take you seriously? --- space reserved for when I have more time --

It says a lot to me that you are referencing AlphaGo which was beaten 100-0 by AlphaGo Zero. That was 2015, mate, an eternity in DL-time. Either you didn't know about AlphaGo Zero (which as a researcher in the AI space would be weird) or you omitted it because it doesn't fit with your beliefs. Either way, not including AlphaGo Zero is like debating navies where one side is thinking wooden sailing ships and the other is thinking nuclear submarines and GPS missiles. The only thing in common is the laws of physics. Let's just pause for a moment and put this into its just context. AlphaGo played 30 million Go matches over many months to learn how to beat the world's best human Go player 4 matches to 1 in March, 2016. AlphaGo Zero taught itself how to play without any human intervention and beat AlphaGo 100-0. It learned by playing 4.9 million matches against itself in 3 days, accumulating knowledge that would take a human thousands of years to learn. Surely a BletchleyGeek appreciates the colossal difference. I don't know who is giving you this information. I'm impressed that you know this - I don't think even the top VCs in this space are aware of your above statement. However, I think you are conflating time-to-market and regulation compliance with technology. AV startups/projects are under immense pressure to deliver results in a given timeline, and there are enormous risks both to capital and more importantly to human lives. I would also be deliberately hiring people with cognitive biases. It's one way to get hybrid models to work. That's a self-disqualifying statement. I know it's en vogue to poke fun at the US military and Silicon Valley in the same breath, but unless you have a very clear idea of what the Pentagon have awarded to Palantir, and on the basis of what due diligence, you may wish to refrain from such a sweeping statement, even if you believe you are sweeping up "pure" bull****. This would be true if this were the stance. But it isn't. Nature observes the laws of physics; AI observes some basic rule set that is compatible with the desired outcome, e.g. AlphaGo Zero used the existing Go board and pieces as its basic rule set. But that was it. No human training, no guidelines, no previous frameworks. So what I am objecting to (that comes across to you as scornful, disingenuous, etc) is the outdated and overused term "AI" insofar as there is no real intelligence happening in the vast majority of games. Unless you can show me how the algorithms are self-adapting to my aggressive or passive style of play, and will remember for the next battle how I conducted a particular set of moves in a given situation, I maintain that the use of the term "AI" in this context is outmoded. If you want to get sentimental you can call it "GOFAI." I would be thrilled if the algos are indeed self-adapting and can tell the difference between me playing and a friend playing on the same PC. Please enlighten me if they are. Not-so-widely hyped in the grand scheme of things (maybe a short mention on the nightly news), but justifiably hyped. This argument is now short-circuited by the AlphaGo Zero omissions, but the fundamental problem with your mental model is that DL in AlphaStar is contributing 99.999% of the heavy lifting. It would be like saying a Prius with a self-parking feature and a solar powered autonomous aircraft are both hybrids of old and new technology. That is objectively true, like it is objectively true that the HMS Victory and HMS Dreadnought (nuclear sub) are both RN warships. Please show that comment. I do not see any comments blasting BFC or anyone else. I do see a comment that is tongue and cheek about "excuses" and "AI" which is clearly contextualised by the sentence that follows it: If encouragement is blasting, perhaps you had a rough schooling. This seems a bit like blasting the game developer for not using AI. I seem to be working pro bono right now updating your DL knowledge, so I guess I had better not set a precedent. Page 17 of the manual: select a Main and Auxiliary section. Page 24 of the manual: Set Priority Targets. Page 24 of the manual: set Commander Observed Fire The Steam forums are full of Guides and tips. Here is a comprehensive YouTube tutorial: https://www.youtube.com/watch?v=wwwacMv-IjQ&t=1048s And a list of cognitive biases you may wish to familiarise yourself with: https://en.wikipedia.org/wiki/List_of_cognitive_biases

I used to find the Tactical Battle UI annoying - now I find that it works really well with a minimum of clicking. The order wheel is pretty intuitive and allows for a lot of permutations. I can give an order to covert advance in tactical column with tight spacing and then queue up an order to spread out into a line with medium spacing and attack using smoke. That's not too clunky in my opinion. This is an unfortunate pathfinding problem if you issue multiple March orders to units that use the same roadway, and I just consider it one of those things I have to micro. I set my March orders to move by road that ends at a bridge, then I micro the bridge as I would have to do anyway in CM. You can issue orders in Graviteam to un-jam a situation. Unless you give a series of expensive "Reverse" commands you can un-jam most situations without exhausting command points. Why are you under the impression that you "have no such control"? Keep in mind that if you play in WEGO, which is what many people play in CM, you have no option to intervene while your units bumble around for 60s. This is pretty much the bare minimum timeframe to learn especially if you have extensive CM experience. For sure they need a better UI and tutorials, and I hope they develop them. This is generally true, however also understandable as objectives are mostly important things to take and hold. If the "AI" guessed where your men were, you would wonder if it was cheating. It's a catch-22. This is kind of unavoidable, though. The enemy has to attack through the open in many cases - if they bunched up along some cover you would switch to shelling the obvious concealed routes and you would wonder why they don't vary it up. If you play Under The Cruel Star (DLC), for example, the Soviets attack through the open but they also have spotters who will drop rounds on you as soon as you reveal your positions. Perhaps you were lucky enough to always kill off their spotters or their units didn't have any assigned. I find they also move through gullies and make extensive use of defilade. Covert Move is pretty powerful and the "AI" does use it. Interesting - perhaps I started playing when this behavior was already updated in patches. I sometimes do have the football kickoff encounters, but generally those are in wooded areas or when both units are concealed to each other by a rise in the ground and suddenly get revealed. I never see opposing platoons charging at each other and not reacting over open ground with good visibility. It is true that sometimes things like the LMG gunners running forward make absolutely no sense and that does frustrate me, but it's thankfully quite rare. As for heavy weapons attracting fire, they also tend to give their positions away fairly easily, but I have not experienced receiving small arms fire from 700-1000m away unless it's from HMG fire. But in general you want an MG-42 HMG to be further back than 700m as you are losing the standoff advantage that comes with the telescopic sight. I tend to place them 1000-1500m back from where I expect the enemy to be. If that's not possible in the terrain, I give them a lot of cover. They are on a big tripod and if they aren't dug in, will be easy to hit. They were the #1 target for Soviet snipers, even above officers. Again I'm not sure why you have this impression. I intervene to give units specific facings and fire arcs or new movement commands all the time. I just don't over-rely on it. This is always the "excuse" from developers who struggle to create a challenge without resorting to "AI" cheats. It is immensely challenging to write algos that will challenge a human player, which is why we do have to encourage devs to train neural networks. If I were Battlefront I would record every H2H match and use it for training data to create a battle AI from scratch. They have a big advantage in this regard and I bet Eugen is already doing it with Steel Division. I do not that this is the 3rd time you've mentioned you don't have the ability to intervene, when you actually do. Knowing the engine's weaknesses, I plan the routing accordingly and conserve command points for intervention. It's like having to play MP. I know it's going to sound a bit off-putting to some, but 25 hours is barely enough time to experience 10 properly planned battles in addition to learning the UI, mechanics, and Operations system. This is a testament to the learning curve, of course, and in the Raid DLC you do have a large number of obstacles for vehicles early on, which is why you ran into the game's biggest weakness if you expected more automation. I think what would make CM players embrace it more would be a way of lowering the Command Point cost of intervening. Having said that, I have often had situations where due to a Command Point penalty (loss of an officer), I couldn't intervene manually and the Computer figured it out, or made an even better move than I was thinking of doing. I have a laundry list of features I want to port from one game to the other, essentially creating an amalgam of both, and add the infantry movement of a FPS. I approach both with different desires - for example I enjoy the urban combat in CM which is very clunky in Graviteam. I enjoy being able to set up a tactical engagement in ways that "real soldiers would." That, in essence is what separates the titles for me. If I could have Graviteam's scale, AFV and combat physics and production values and with CM's troop behaviour, I wouldn't do much else all day long.

Spending more time on mods than playing the game defines me fairly well. Which is one reason why I plan to start a gaming dev company when I sell / exit my current business. I want to create a series of SDKs based around AI frameworks so modders can focus on training soldiers, units and crews using data sets from a few hundred thousand gameplay recordings from Twitch and YouTube. This is the modding of the future and will make it possible to have 1,000 vs 1,000 battalion level battles with realistic behavioural models and no need for manual scripting or skinning. Back to work...

Define "extremely clunky games." Which aspects are clunky? How many hours have you played GT? I had a series of expectations coming into GT from CM and other titles, and once I learned the UI and C&C system and philosophy behind the game design, it grew on me immensely and I appreciate both games in equal measure. It was kind of like going from Gary Grigsby to HOI3. Context: I bought my first CM game in 2004 (Afrika Korps) and my first GT game in 2018 (Tunisia '43). I have over 500h in both "systems" though recently returned to CM after a long time away. I respectfully disagree with your assertions. If you are defending, then 90% of your outcome should be based on where you position your men and heavy weapons, and whether the enemy's artillery and heavy weapons fire is accurate and intense enough to destroy key emplacements, and you should always be looking to achieve a men-materiel-and-terrain advantage over your opponent at the operational level, which will lead to asymmetrical force structures and dispositions in engagements. If you are the attacker, then the Graviteam "AI" is overly aggressive on defense, which is indeed frustrating, and that is down to its eagerness to seize victory/key points if it cannot see all of your forces and you happen to be sitting on a V/KP. However, many times I have been surprised when I have crushed what I thought was a pre-emptive (and ill-advised) attack only to find an entire company remained dug-in on the main objective. If you regularly win operations hands-down in GT, you are one of the very few who do. The "AI" in GT rarely applies the same approach in the same exact battle (you can test it yourself by restarting). By contrast, I cannot help but feel that CM is scripted to a point where the AI will try small variations on the same thing every time based on the way the decision tree of interconnected conditions unfolds. The easiest way I have found to test that is to play the smallest battle. This is where CM should truly excel due to the emphasis on small unit tactics, but the TacAI will not deviate from its hand-made plan. All the human has to do is re-position one AT gun and the "AI" struggles to counter that. Disclaimer: I work in the field of AI and Computer Vision, and calling a series of interdependent if-statements, "AI" is, at best, to apply an antiquated definition, and, at worst, to completely misunderstand what AI actually is. I prefer "Computer Opponent" or "Algorithms." When we eventually do play vs AI, we will be begging for mercy. I do not believe (but would be happy if it were the case that) either game designer collects vast amounts of gameplay data and supplements it with multiple camera angle footage of soldiers in combat, and layers the computer vision data from cameras mounted on AFVs from Syria (opponents need to shoot back) allowing a series of Neural Networks to process and construct autonomous tactical models. The closest I have seen AI in battlefield simulation is the work done by Palantir Technologies and the DCGS-A system they are co-developing. So what we are contrasting here is different approaches to C&C design, scripting and algorithm writing. With any system, for every 100m of combat front, 10 soldiers and 1 AFV you add, the complexities go off the charts. Case in point in CM: do a Quick Battle with 15k force budget and field an armoured battalion on a "Huge" map setting against a recon infantry opponent. The game engine struggles mightily to drive the 40+ tanks across the battlefield in any coherent manner. The front few race ahead while the majority smash into each other in a massive traffic jam. In Graviteam, the same command ("Move") produces a much more organised outcome, on a 3km x 3km map, for a variety of reasons, and you do not need to micro-manage AFVs to take roads. Primarily it is because Graviteam started life as an AFV simulator company and CM apparently started life simulating tabletop games like Crossfire (although thankfully the dice rolls have all but disappeared and I believe all projectiles are simulated and not probabilistic). In Graviteam, I am able to split off a 2-man squad with one click (Alt-LMB) and have them covertly move to the crest of a ridge. I can lay down area ("observed") fire by moving the platoon leader to a position from which he can observe the target area and selecting which squads to switch to "AI Fire Control." I can also split up squads into fire teams, hotkey them, then quickly assign individual enemy soldiers as priority targets, as if I were Cpt. Winters in the Crossroads episode of Band of Brothers. So that fine control is present. However, that is not the heart of the game. CM is excellent for small unit tactical simulation on a relatively small battlefield and it has no peers when it comes to the degree of fine control that a human player can apply to squads and fire teams. To me it feels like playing Crossfire on a computer, but without the dice. Its WEGO and H2H are fantastic, although I would rather have a replay system and train AlphaStar to play it rather than rely on other humans to have the time, energy and skill to play me. Graviteam (Mius Front / Tank Warfare) is a 3km x 3km battalion level simulation that actively discourages micromanagement and excels at recreating the relatively hands-off experience that a Captain or Major would have experienced in WW2. It is lacking replay/WEGO and H2H, and it needs better documentation. I think CM players should absolutely play Graviteam, and vice versa, take the time to understand them, and play them for their strengths, not their weaknesses. They're both worth your time.

Thanks - I read that Mord's UI faces mod is not required if I already have Gustav Line. Is that still the case?

This makes sense. Strangely, though, many other modded vehicles appear in the "Firing Range" scenario, just the M4A1 Mids do not.

Yes - the M4A1 mods in question have the CMFI tag and are here: http://cmmodsiii.greenasjade.net/?p=873 http://cmmodsiii.greenasjade.net/?p=869 I have added both of these folders to /Data and to /Z. Neither triggers the modded decals in the Tutorial mission #2. It's not a game-breaker I am just trying to understand which M4A1 "Mid" unit is in the TOE and how many M4A1 "Mid" models there are if this particular one in the mission does not match either of the mods linked above.