Master of Magic 2 Wishlist For Slitherine

[haphazard1]

Agree with Blake00 that many of the possible "major" variations of MoM already exist in other games/series. I have played most of them, and enjoyed some of them quite a lot. But they are not MoM.

The only reason a company buys an existing IP and puts it on something totally new is to try to lure in fans of the old IP to try their game. If the new product is too different from the original, a lot of those fans will feel cheated/misled. The company may say "It worked, we got their money, who cares that they are unhappy?" but that harms the company's reputation for future games. Hopefully Slytherine is looking to make a solid game and build longer-term appeal, and not just for a quick cash grab.

[Arnuz]

Guys come back to earth, no one's ever going to "finish developing the old mom".

Just take a peek at slitherine's games. Panzer corps: dumbed down pg. Polaris: moo with effing star lanes. Pandora: aow on a new planet, but only three races. Whatever makes you think that they would carefully study the balance of 200 spells and 20 races on two planes?

We'll be lucky if it isn't 3d. Very lucky. I'm not even hoping. If "lol owning the rights and putting the name on it won't make it a real sequel" well, then it won't be.

I also think that decent AI is a must. Unless it takes only a short amount of time to play the full game -- a few hours at the very most -- getting a group together who can consistently play to keep a multi-player game going is tough. So most players are going to be playing single player, and that means AI opponents who can actually play the game and provide a decent challenge. The game mechanics must take the need for effective AI into account -- anything the AI can not handle needs to be removed or modified, no matter how cool the designer thinks it is. (Hello, 1UPT Civ games, I'm looking at you.)

Testing, lots of testing, is very important. Including AI testing -- something like Sullla's Civ AI Survivor would be a very good idea. Make sure your AIs can expand, attack and conquer a rival, defend themselves from an attack, and actually reach a victory condition. Preferably all the victory conditions which exist in the game, if there are more than one. AIs should actually use all the major features in the game, not ignore some of them. (This is usually more of a problem with expansions, but sometimes occurs in initial releases.)

How many examples of this have you seen recently where it didn't depend on dumbing down the game? Moo4 I'm looking at you...

Don't get me wrong, I subscribe to 100% of your comments wholeheartedly. But, as things are, this'll remain wishful thinking, unless they drop the idea of building in-house AI. Leave that to AI experts and focus on what you're good at: making the game. Not accepting that someone else can be better at ai would just be the "not invented here" antipattern.

Seriously, this level of artificial intelligence doesn't exist yet. The day that thing becomes reality, will be the day I stay in bed all day and let the AI write my forum posts :D

If you stop for a moment developing in assembler you might notice that the cloud has happened. That's what I'm talking about in the post:
- make mp mom
- make ai agent using deepmind
- use the money saved by not developing bad AI to code more races and spells into the game
- make agents fight themselves in cloud instances
- do fight/rebalance cycles till race/realm balance is achieved

... And laugh at nerds complaining about no more invincible swordsmen with life buffs. Or no more running around to prevent a city occupation. Or....

[Seravy]

Just take a peek at slitherine's games. Panzer corps: dumbed down pg. Polaris: moo with effing star lanes. Pandora: aow on a new planet, but only three races. Whatever makes you think that they would carefully study the balance of 200 spells and 20 races on two planes?

I don't THINK they will do that. But if they don't, I also don't CARE about whatever they do. There are dozens of modern games around already, even in the 4x genre, more than I could play even if I did nothing else for the rest of my life. My "to try" list is full of them. It needs to be much more than that to be relevant. More precisely, 4X games have high replay value. I have no reason to play any of them if they aren't equal to or better than MoM/CoM, I can just keep playing MoM and CoM instead. This isn't like an RPG where you clear it once and need another one. A good 4X game lasts your entire life or close to it.
Honestly, if I expected any game company to make a good enough game, I wouldn't bother modding MoM myself. But we can at least let them know what they're getting themselves into...

If you stop for a moment developing in assembler you might notice that the cloud has happened. That's what I'm talking about in the post:

I understand what you're saying but I don't think deepmind, or any other currently existing AI is at the level to play a complex 4X game well. If that was a thing, it'd be in every single newspaper and we'd hear about it winning the Nobel Prize. You can't demand they use future technology that doesn't exist yet. You're the one that needs to come back to Earth here. Cloud is nice but that only means the supercomputer running the AI doesn't have to be in your home to play. It still has to exist somewhere.

[haphazard1]

As a player, I do not really care if a game's AI is running in the cloud or functions because of a mechnical engine carved from stone blocks. It just needs to work. And it will be the game dev building the AI, because who else would? Cloud-based or not, an AI capable of playing a middling-complex 4X strategy game is not a simple thing to create. Only the game developer is going to have the incentive to put in the time and money needed to build it (and test it and balance it) for a new game.

Once a game exists and is popular, you might see fan-sourced and crowd-sourced efforts to improve such an AI. This has happened for a number of games. But that fan base does not exist until after a good game has been created. Maybe someday we will see companies that do nothing but produce game AIs, who get hired as partners/third party help by game devs who focus on just creating the games. But we do not seem to be there yet. (Hmmm, business opportunity....)

[Blake00]

Polaris: moo with effing star lanes. Pandora: aow on a new planet, but only three races.

Oooh I didn't know about either of those games from Slitherine. They don't rate highly on steam but they look okay and many of the reviews (even the bad ones) all seem to talk about how hardcore the AI is so that's promising. Yeah they're not amazing games but they actually give me comfort as it shows years of experience with 4x strategy gaming. Hopefully they take the good stuff from these and learn from the bad stuff people didn't like and put them to good use with their MoM project. Although it looks like different developers worked on those games and Slitherine just published them, so I was wonder who they will put on MoM.

PS is it just me or is Pandora using the same world engine as Warlock? That is so darn similar.

[Arnuz]

Blake, it's easier to make decent traditional AI if they dumb the game down enough. Pandora and the other games from Slitherine are examples of this, as is moo4. That's though, exactly the opposite of what we're asking for: the fully fledged mom experience, the "remaining 80% that wasn't coded when the game was published at 20%" to paraphrase Seravy...

I understand what you're saying but I don't think deepmind, or any other currently existing AI is at the level to play a complex 4X game well.

Cloud-based or not, an AI capable of playing a middling-complex 4X strategy game is not a simple thing to create.

Once a game exists and is popular, you might see fan-sourced and crowd-sourced efforts to improve such an AI. This has happened for a number of games. But that fan base does not exist until after a good game has been created. Maybe someday we will see companies that do nothing but produce game AIs, who get hired as partners/third party help by game devs who focus on just creating the games. But we do not seem to be there yet. (Hmmm, business opportunity....)

Here you guys are completely mistaken, perhaps you don't know recent AI development well enough. Read up about deep learning a little bit.

There's no "developing the AI" anymore, that's a thing of the past. You just develop the agent that plays in multiplayer, then - after a "salting" of a short supervised learning (that for the human might equate to, say, following the tutorial) you just set up the agents to play against each other, unsupervised, and improve based on the statistics.
The advantage of the cloud is that now you can make the AIs learn unsupervised in a fraction of the time it took before: https://www.theverge.com/2017/10/18/1649...elf-taught
Months to teach the AI to beat the players in supervised mode... And then, the unsupervised learning AI beats the supervised one (the one that beat humans) after 3 fucking days of playing against itself. In 3 days it "had accumulated the equivalent of 100 years of playtime" IIRC.

But your bigger mistake is about complexity. Go is the simplest rulest that exists - 1 piece, 2 or 3 rules - far simpler than chess and other games. But you can put a piece on any unoccupied space, and as a result, Go has much wider possible move search space - estimated at 10^170, or more than the square of the number of atoms in the universe. This is what matters, not the number of units, spells and races - sure that complicates it, but linearly. So in chess - which has I think 6 times the pieces and moves of go - the complexity is around 10^50.

Now, StarCraft is obviously more complicated than Go. But your assertion that turn-based 4x games are even more complicated is bizzarre. Starcraft is real time, and scientists have assessed its move space to ~10^8.000.000 choices based on the average length of a pro competition. (not bad of an improvement from 10^170 huh since 2017 when Go was beaten!)
I can't think of any assessment of move search spaces for 4x games but I'll go out on a limb and say that a turn-based game is IMHO much less wide in terms of search and possibly deeper - fewer choices (realm time vs turns) but among more possibilities (spells, units, realms, although actually starcraft has more resource types) - but width is what makes the space swell. Before you answer think again of chess vs go: chess is more complicated, but go has the bigger move search space. I know something of AI and I am nevertheless first to admit that I cannot be sure about this, but intuitively I am ready to bet that 4x games have a much MUCH smaller search space. (hint: map size doesn't matter if you abstract well the moves: what matters is time/turns to objective as I've tried to convey to you several times Seravy, without much success)

So to conclude: haphazard, your idea that "Once a game exists and is popular, you might see fan-sourced and crowd-sourced efforts to improve such an AI" is really off. Nowadays coming up with these algorithms is the realm of deep learning scientists, not of game devs or fans (much less improving them). And if you really think that then you're in for a real surprise when you try your wits against the machine. It's not going to be anything like in the movies where generally human creativity carries the day. Both Kasparov and the StarCraft pros when interviewed after the matches said something similar: Kasparov surrendered after being suprised, and the starcraft pros had to go up against never before seen strategies.

[Seravy]

Interesting. In theory real time does mean there are more "turns" to make choices, yes. But in games like Starcraft, you can't really move all your stuff all the time. Once you're producing something in a building, that building can't do anything for a very long amount of time. And most of the time you are producing something. If you aren't that means you have no resources to do so (either because you're saving them for a bigger thing, or just don't have enough at all). So there aren't all that many decisions, if I say you get to pick one thing to build out of 10-20 available options per 10-20 seconds, that's about reasonable. So complexity is 10^(time/20) for buildings. Moving the units is a different story but you admitted that part is limited (AI can't click more often than the human) so they can't micromanage it to the milisecond and pixel detail.
In MoM, you own 20 cities all of which produce independently without using your resources. Even assuming you only get to pick one of 10 choices and only 5 cities finish production a give turn, that still is (10^5)^(turns). That's 4 magnitudes higher, per turn played, just on the buildings. And each turn each city is also allowed to set farmers, taxes, and we haven't moved units that, which when engaged in battle, start their own minigame of 25 turns of 9 units, pretty much a chess game of its own. With spellcasting as a bonus. We also haven't used the diplomacy subsystem, or moved equipment on our hero units. Or used an Overland spell (create artifact by itself has over a million possible ways to make the artifact. It dwarfs the complexity of all other spells by itself.)  So I estimate the base complexity of a MoM-like 4X per unit of time played to be at least 10 magnitudes higher. As each unit of time is then multiplied, overall the complexity is unimagineably larger.

So yes, I completely agree with what you say about the AI but you are underestimating the complexity of 4X games. I don't think it's viable (yet, maybe in another 10 years...). Another issue is, using a cloud means internet connection required. And someone has to pay for the AI fees. Which means the game has to have a subscription fee and/or will cease to function as soon as the company releasing it decides to discontinue support.

[Arnuz]

- Each building in StarCraft can do different products, so it can be compared to a simpler city - already with some basic buildings given that the cities can only do spearmen at first. When the city is more complicated that still only raises the search space linearly, and usually that's only the capital or the first few cities.
- A million? Peanuts. And when you multiply by a million, or a billion, again that's only a linear increase. Not that it's actually a high number if you think about what really matters to heroes, but even if it were... Still peanuts.
- diplomacy, lol, it's either status quo or not (change to peace/war). Nothing else matters and neural networks are pretty good at discarding what doesn't matter.
- non obvious outcome tactical battles depend mostly on the Wizard spells, which is again the perfect application of neural networks: space segmentation, and elimination of the least favourable situation/choice
- the big cost of the cloud is for training - "100 years of effective play time" in 3 days for Go. The cost for running the player AI will be in cents per person per year. I'm pretty sure Google will agree to make it free for players as long as the company does the training with them.

I'm not underestimating 4x, I'm pretty familiar with both types of game. In the end it depends on the abstraction that is done to simplify the move search space. As previously mentioned I'm not 100% sure that 4x is simpler. I give that a 80% git feeling likelihood, but even if in wrong, either it's in the same whereabouts, so it can be done like Warcraft, or it's at worse somewhat more complicated, then it needs like what, 1 month of wait?... Don't forget:
1. the jump in solved complexity in the 2 years after Go: from 10^170 to 10^8000000
2. This last January IBM put the first quantum bit chip on the market..... We might be in for an acceleration of Moore's law in the next few years

There's absolutely no way that trying to do old school AI could be the best way nowadays to ensure that slitherine's spending money where it should - on the game design.

[Seravy]

I guess we'll know who is right in a few years. In the end, we are both guessing on the complexity. If it is possible to use that type of AI for 4X already, someone will sooner or later do it and then it'll be obvious.

[haphazard1]

I think a modern 4x game is rather more complex than chess. Yes, the space of possible moves/games in chess is large. But chess is a pretty simple game in a lot of ways:

- Playing area is small (64 spaces), never changes, and each space has no properties other than relative position
- Very limited number of different types of pieces, and these types do not change. One type may be replaced by another if a pawn is promoted, but knights never start moving an extra space because you researched better roads or whatever.
- Start of the game is not variable (although I guess there is handicapping like removing a piece or two for a stronger player to make an even match, this is pretty minor as variations go.)
- Game structure is very simple with exactly two players who alternate turns, with each player performing exactly one action per turn. The other player and his pieces can not affect the acting player's move in any way. (I suppose en passant capture is a minor exception.)
- The outcome of every player action is entirely deterministic
- No entities in the game other than pieces (no cities, etc.)
- The definition of victory is very specific. (Not technically a single victory condition as there can be stalemates or unable to make a valid move results, but 99+% of all non-resigned games end in one way.)

I think 4X game AI will prove plenty complex. I hope someone does come up with a good solution, because I would love to play better AI opponents. But I think it is going to be a bit more complicated than writing an agent and running a few thousand hours of simulated game time.

Edit: Forgot a huge one. Chess ia a game where all players have perfect knowledge.