"Rome 2 launched with a largely positive response" Can you elaborate what you mean by this? Critic wise it got pretty reasonable scores but the player base was furious with non reactant AI and a bunch of gameplay issues, even so that it still taints Rome 2's name to this day even with all the fixes implemented.
Yes I meant in terms of critics rather than player base. I clearly didn't express this well enough as I know the fans were NOT happy with it on launch, sorry about that. It's actually a point of focus at the beginning of the next video in that they had to patch it a ton between Rome II and Attila for it to work properly.
Dr. Stephen Strange: I went forward in time... to view alternate futures. To see all the possible outcomes of the coming conflict. Peter Quill: How many did you see? Dr. Stephen Strange: Fourteen million six hundred and five. Tony Stark: How many did we win? Dr. Stephen Strange: ...One.
Love the series and the lectures were awesome as well, keep it up! On another note, would love it if you looked at the differences between Total War and Europa Universalis IV, if the AI of the latter is any special. I always felt it more consistent and more intelligent.
Wait so did the battle AI stay more or less the same from Shogun 2? Curious since at launch I remember that most of the dumb decisions that I saw the AI make were in battles, in the campaign it worked decently enough.
I'm not 100% sure about when/how the battle AI shifted around this period. I have a much clearer view of how the battle AI is modelled in the likes of Warhammer. Hence I don't get around to talking about the battle AI again until part 5.
Like others have said. Rome 2 was a regression in AI for the total war franchise. The shear amount of exploits and cheesing a player could do is well documented across YT. It had it's moments but I think CA were too focused on trying something different to really address the problems they caused.
The problem I have with this video is very similar to the problem I had with the game at launch. You didn't playtest your own assumptions. I'm glad the theory is interesting, but that AI fails to solve the most basic problems correctly. Even if such problems aren't solved properly in a relatively low number of occurrences, if these failures are so bad that it is immersion breaking and observable to the human player, then I just don't care that it can handle a large volume of problems in a very efficient way because it's irrelevant. The challenge was to create the AI of a STRATEGY GAME by using algorithms that others have used to solve other problems. We already know that the algorithms work in certain situations and that it has some limitations. The R2 devs had to adapt their usage of those tools to the particular task at hand. In the end, the AI can't realise when it is making completely terrible decisions from a strategic perspective. That's kind of a biggie. It's a strategy game, it has to minimally do the most basic things properly. Period. This is particularly striking that you fail to touch on this point and the game's devs missed it too. It's not like they're bound from handing the corner cases where the algorithm fails with other means, even static decision making would have been preferable to the final product. The discussion needs to happen, otherwise people will just repeat these mistakes.
It always Astonishing how dumb the total war player base is you cannot disregard what was said in this video about the complexities in the Ai using anecdotal evidence ignoramus
One interesting thing you seem to point about R2's AI system is its scalability (somewhat). You can let the algorithm run longer to find nodes in the tree, giving it a better chance at finding a better one. Combined with the possibility of looking farther into the future than 1 turn and greater CPU horsepower in the future, it seems to me like this iteration of the AI could last a while!
Absolutely. MCTS is capable of near-optimal play in much smaller problem spaces (e.g. old Atari games like Pac-Man). The more complex (and random) a given problem is, it requires more evaluations (and in turn resources) in order to make smart decisions. There has even been research in how best to customise AI difficulty levels by tweaking the parameters of the playouts by simply giving more/less time to think on a given frame.
Hey, great video! I must say, you are bizarelly knowledgable on the specifics of Total War's campaign AI, even down to particular project-specific nomenclature. Have you talked with campaign AI programmers from the Creative Assembly? My only comment is: does MTCS itself come under the domain of machine learning? It's a graph exploration technique, where the AI does not learn from past decisions, only from the scoring within branches of the tree. AlphaZero apparently combined this with an ANN, which would make it machine learning, but the MCTS algo itself is not as far as I'm aware.
Thanks for the comment. I have met a handful of Total War AI devs over the years, but these videos were done without their assistance. Although one ex-CA developer did review these videos for me prior to release to check authenticity. To answer your question: Yes you're right that MCTS is a graph tree search algorithm, but it is still reliant on principles of Monte Carlo learning (a form of reinforcement learning). The system is learning an optimal behaviour for a given scenario by continually experiencing (or simulating) outcomes of a given Markov Decision Process it has no defined model of. The reason MCTS is so powerful, is that prior to now the structure of this experiential learning process wasn't so strictly defined. It takes good practices from graph exploration (such as MinMax and Alpha-Beta pruning) and applies that atop an existing machine learning model. It just wasn't possible 10/20 years ago because the simulation phase is rather CPU intensive. I think I need to do an episode on this at some point! 😂
Man when this gane Launched every battle was a massaker. The AI just targeted the Flag Area and Moved all units to this point. My Soldiers just stand there in their way and the enemys got killed without even fighting because they marched through my soldiers. Or Playing as Sparta was a pain in the ass. Knossos decleared war and landed with 6 Full Stack Armies in front is sparta. I couldnt even recruit 3 Troops in the same time. And after i killed all of them the next wave just landed in front of Sparta. Unplayable
somehow i feel like the AI is slowly but surely learning from my gameplays because its adapting its armies against me and they get harder every game D: Rome 2
Rome 2 was such a buggy mess when it first came out. I gave up on it after a few hours. (laying siege and 5 units of mine getting stuck on a open gate not moving ugh). Came back to it over a year later and it was a different story. Fun game once it was patched to be playable.
i played since the first total war, shogun 1 the one that totaly sucked me in was rome 1 and the one that totaly drive me away was rome 2. Rome 2 was the so bad at launch, so so bad that i quit the rome 2 and never come back to it. The worst of the worst AI and bugs so huge and bad that made the game unplayebel. I went back to medieval 2/kindoms and stay there and never bought a total war game since.
Rome 2 Campaign AI is supposed to do all of that? nothing of this can be seen in the actual game, the AI is completly braindead sitting at their capitals with 1 full stack of tier 1 units. How can you mess up an interesting AI as described in this video so bad CA....
The Ai of shogun and Rome 1 was the furthest thing from being smart also its not that simple as rome 2 is far far far more complex then both of those games ,it's the reason why the first 2 game had "better " Ai then rome 1 and medieval 2 because they had less to deal with and the games were more simplistic
"Rome 2 launched with a largely positive response" Can you elaborate what you mean by this? Critic wise it got pretty reasonable scores but the player base was furious with non reactant AI and a bunch of gameplay issues, even so that it still taints Rome 2's name to this day even with all the fixes implemented.
Yes I meant in terms of critics rather than player base. I clearly didn't express this well enough as I know the fans were NOT happy with it on launch, sorry about that. It's actually a point of focus at the beginning of the next video in that they had to patch it a ton between Rome II and Attila for it to work properly.
No need to apologise it's a great video! I was just curious.
I was wondering too lol. Part of the disaster even was that the ai often failed to work, never mind do the things it was advertised to do.
Dr. Stephen Strange: I went forward in time... to view alternate futures. To see all the possible outcomes of the coming conflict.
Peter Quill: How many did you see?
Dr. Stephen Strange: Fourteen million six hundred and five.
Tony Stark: How many did we win?
Dr. Stephen Strange: ...One.
Love the series and the lectures were awesome as well, keep it up! On another note, would love it if you looked at the differences between Total War and Europa Universalis IV, if the AI of the latter is any special. I always felt it more consistent and more intelligent.
Thanks. I''ll add Europa Universalis to the to-do list and see what I can find.
Wow I'm here early. I've always found the Rome 2 AI interesting because it was mostly brain dead but sometimes could pull off great moves.
it should was better than Empire
Love the Total War franchise and love the information and insight you're giving. Glad I stumbled on the channel!
Wait so did the battle AI stay more or less the same from Shogun 2? Curious since at launch I remember that most of the dumb decisions that I saw the AI make were in battles, in the campaign it worked decently enough.
I'm not 100% sure about when/how the battle AI shifted around this period. I have a much clearer view of how the battle AI is modelled in the likes of Warhammer. Hence I don't get around to talking about the battle AI again until part 5.
Aipe97 shogun 2 ai were smart and they use different tactics as a player does
This was a little more useful than the #2 Total War video. It still leaves a lot unanswered, but at least it made me think. Thanks.
Rome 2 lunch to a largely positive response??!!! Have you not see the shit storm at lunch???!!!
Nah, I had a big breakfast.
Like others have said. Rome 2 was a regression in AI for the total war franchise. The shear amount of exploits and cheesing a player could do is well documented across YT. It had it's moments but I think CA were too focused on trying something different to really address the problems they caused.
The problem I have with this video is very similar to the problem I had with the game at launch. You didn't playtest your own assumptions. I'm glad the theory is interesting, but that AI fails to solve the most basic problems correctly. Even if such problems aren't solved properly in a relatively low number of occurrences, if these failures are so bad that it is immersion breaking and observable to the human player, then I just don't care that it can handle a large volume of problems in a very efficient way because it's irrelevant.
The challenge was to create the AI of a STRATEGY GAME by using algorithms that others have used to solve other problems. We already know that the algorithms work in certain situations and that it has some limitations. The R2 devs had to adapt their usage of those tools to the particular task at hand. In the end, the AI can't realise when it is making completely terrible decisions from a strategic perspective. That's kind of a biggie. It's a strategy game, it has to minimally do the most basic things properly. Period.
This is particularly striking that you fail to touch on this point and the game's devs missed it too. It's not like they're bound from handing the corner cases where the algorithm fails with other means, even static decision making would have been preferable to the final product. The discussion needs to happen, otherwise people will just repeat these mistakes.
How is the ai now 3 years later?
It always Astonishing how dumb the total war player base is you cannot disregard what was said in this video about the complexities in the Ai using anecdotal evidence ignoramus
You failed to mention that Rome 2 was just about unplayable at launch, although it works pretty well now, especially with the DEI mod.
I do mention it needed work with some patching. Though I will be elaborating on the issues it had at the start of part 4 next week.
I cannot say I think you did.
such great content! Thanks!!!
One interesting thing you seem to point about R2's AI system is its scalability (somewhat). You can let the algorithm run longer to find nodes in the tree, giving it a better chance at finding a better one.
Combined with the possibility of looking farther into the future than 1 turn and greater CPU horsepower in the future, it seems to me like this iteration of the AI could last a while!
Absolutely. MCTS is capable of near-optimal play in much smaller problem spaces (e.g. old Atari games like Pac-Man). The more complex (and random) a given problem is, it requires more evaluations (and in turn resources) in order to make smart decisions. There has even been research in how best to customise AI difficulty levels by tweaking the parameters of the playouts by simply giving more/less time to think on a given frame.
Hey, great video! I must say, you are bizarelly knowledgable on the specifics of Total War's campaign AI, even down to particular project-specific nomenclature. Have you talked with campaign AI programmers from the Creative Assembly?
My only comment is: does MTCS itself come under the domain of machine learning? It's a graph exploration technique, where the AI does not learn from past decisions, only from the scoring within branches of the tree. AlphaZero apparently combined this with an ANN, which would make it machine learning, but the MCTS algo itself is not as far as I'm aware.
Thanks for the comment. I have met a handful of Total War AI devs over the years, but these videos were done without their assistance. Although one ex-CA developer did review these videos for me prior to release to check authenticity.
To answer your question: Yes you're right that MCTS is a graph tree search algorithm, but it is still reliant on principles of Monte Carlo learning (a form of reinforcement learning). The system is learning an optimal behaviour for a given scenario by continually experiencing (or simulating) outcomes of a given Markov Decision Process it has no defined model of. The reason MCTS is so powerful, is that prior to now the structure of this experiential learning process wasn't so strictly defined. It takes good practices from graph exploration (such as MinMax and Alpha-Beta pruning) and applies that atop an existing machine learning model. It just wasn't possible 10/20 years ago because the simulation phase is rather CPU intensive.
I think I need to do an episode on this at some point! 😂
Man when this gane Launched every battle was a massaker. The AI just targeted the Flag Area and Moved all units to this point. My Soldiers just stand there in their way and the enemys got killed without even fighting because they marched through my soldiers.
Or Playing as Sparta was a pain in the ass. Knossos decleared war and landed with 6 Full Stack Armies in front is sparta. I couldnt even recruit 3 Troops in the same time. And after i killed all of them the next wave just landed in front of Sparta. Unplayable
Your opinion on copy generator?
somehow i feel like the AI is slowly but surely learning from my gameplays because its adapting its armies against me and they get harder every game D: Rome 2
I never really considered mcts a reinforcement learning algorithm
Rome 2 was such a buggy mess when it first came out. I gave up on it after a few hours. (laying siege and 5 units of mine getting stuck on a open gate not moving ugh). Came back to it over a year later and it was a different story. Fun game once it was patched to be playable.
TheVargr do you prefer it now to Atilla? Can't decide which one to pick... And definitely not in a mood to possibly do Steam refund for either...:3
All of this and still not a single TW game with good AI.
Hey Tommy - great videos but for me the music is way too loud and bombastic
that game has an AI?
i played since the first total war, shogun 1 the one that totaly sucked me in was rome 1 and the one that totaly drive me away was rome 2.
Rome 2 was the so bad at launch, so so bad that i quit the rome 2 and never come back to it.
The worst of the worst AI and bugs so huge and bad that made the game unplayebel.
I went back to medieval 2/kindoms and stay there and never bought a total war game since.
they fixed most of the major issues. i played it, it's pretty good now. At least i enjoyed it
Rome 2 Campaign AI is supposed to do all of that? nothing of this can be seen in the actual game, the AI is completly braindead sitting at their capitals with 1 full stack of tier 1 units. How can you mess up an interesting AI as described in this video so bad CA....
And still Rome 2 ai sucks and you can't tell why... Why fucked theory just make the ai as smart as Shogun and Rome total war
The Ai of shogun and Rome 1 was the furthest thing from being smart also its not that simple as rome 2 is far far far more complex then both of those games ,it's the reason why the first 2 game had "better " Ai then rome 1 and medieval 2 because they had less to deal with and the games were more simplistic