Help with testing Fred - Freelands MP Custom AI v0.14.9

Discussion of all aspects of multiplayer development: unit balancing, map development, server development, and so forth.

Moderators: Forum Moderators, Developers

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby Krogen » November 10th, 2017, 4:48 pm

I downloaded the new version and played one.
Yes, Fred is way too defensive. I managed to take all 16 vills and held 14 while Fred still had 10+ units. I took 12 without Fred trying to attack any of my units even once. Wesnoth AI will probably never be a challange for a truly experienced player, but exactly for that reason, i think Fred should be braver. He can still get lucky, that can even the odds a bit. Ofc i don't know anything about programming, so i don't know at all if a bold, but better than Default AI is even feasible.
Anyway, if this defensiveness will be kept then i think Northerners isn't the best faction for it. It would be much harder to break the line if Fred could play Knalgans or Rebels for example.
Naturally it's still fun, it's good that i actually have to plan the battle and not just destroy everything in my path.
"A lion doesn't concern himself with the opinions of the sheep." - Tywin Lannister
User avatar
Krogen
 
Posts: 131
Joined: January 1st, 2013, 3:43 pm

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 11th, 2017, 2:09 am

Hi Krogen — thanks for the comments, good to hear that you got something out of it!

No, Fred is not supposed to stay this defensive, esp. as Northerners, and, in principle he can play other factions already (just that that's disabled for the time being). What I have done over the last few releases is work on the individual pieces: attacks, defensive lines, village grabbing, retreat, ... So that's the tactics layer. The problem (one of the problems?) is that at the moment these pieces are put together and coordinated pretty haphazardly. Working on that part (that's the operations layer I keep talking about) is next on the list. As I said before, until I have done that I don't know how well that will work, but it will definitely be possible to dial up the aggressiveness, and I will do so in a configurable way.

And just to complete the layers, there's really no strategy layer needed for a simple MP map like Freelands, as the strategy is obvious: kill the enemy leader. That might be different if I'll ever adapt Fred (or another AI) to multi-scenario campaigns.

And thanks for another replay, ElderofZion.
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 26th, 2017, 9:42 pm

It's been a while, but I haven't done all that much yet. At the moment, I seem to be getting a few days with some time for Fred, followed by a week or more without any. C'est la vie. In the meantime:

Some (at best) semi-coherent ramblings

  • We all know that playing against the default AI is not how to go about balancing against a human player. No surprise there. What I did not expect was that playing against the default AI is the main reason why Fred is currently so exaggeratedly defensive. However, once I think about it, it makes a lot of sense. Since the default AI just throws everything at you until it is exhausted, sitting back and waiting it out is actually a pretty effective strategy. That's also why the replays you all are providing are so valuable — I simply lack the time to do all the human-player testing I should (and would like to) be doing.
  • I think that the attacks themselves are not the main reason for this defensive play. Sure there's room for improvement there too, but the main problem is not bringing units into positions from where they can attack effectively on the next turn. So that's what I have been concentrating on so far.
  • Working on the operations layer consists of two parts. Assessing the overall situation and coming up with a game plan, but also making use of that information in the individual actions. The former is what the operations layer does, the latter needs to happen in the tactics layer.
  • The good thing is that lots of hooks are already in place in place for these tactics layer adjustments. For example, the defensive holding action has evaluation function contributions for the overall aggressiveness, a "move forward/backward" component, one for moving into (or away from) the most vulnerable areas, one for the ratio of forward vs. counter attack damage at the hold position, etc.
  • In fact, there are definitely too many of these contributions at the moment, but since I do not yet know what works best under what circumstances, I need to keep most for now and play around with them to find a workable solution.
    • As a side note, once the rough balancing is done, fine-adjusting the weights of these functions is probably a good candidate for applying machine learning and genetic algorithm techniques. But I think I need to get it approximately right manually first.
  • The way how I am handling this in practice is by pulling out parameters into a configuration file. Don't pay too much attention to exactly what's in there yet, but the goal is to have a small number of parameters that are somewhat intuitive and can be adjusted easily on an as-needed basis for a given scenario or faction (similar to the aspects of the default AI), with a bunch of secondary parameters that can be use for AI development purposes.
  • We'll have to see what "small number" means in practice. For example, the operations layer now adjusts the base aggression from the config file based on the relative time-of-day adjusted power of the two players. That works well for Northerners against lawful factions, but might not be so great against other chaotic factions. More so, even against different chaotic factions the best overall strategy is not necessarily exactly the same.
  • Ideally, Fred would detect the needed adjustments automatically, but that might be setting the bar too high, esp. once we go beyond the Default Era. So there might be some level of automatic adjustment that results in decent (but equal) style of play against both Northerners and Undead. And in addition there could be secondary parameters for custom behavior tweaks. I don't know yet.
  • In the end I might also add some other parameters for overall playing style. For example, when it's clear that we're losing, we could bunker up to make it last as long as possible, or attack all out to do the most damage possible. Which one is preferable depends on individual preferences, scenario objectives etc. So doing this via an adjustable parameter might be nice. Anyways, that kind of thing is second order at the moment, just thinking ahead. In any case, all the hooks for this are in place already, it's just a question of adjusting the weights.
  • Finally, a note on mistakes that I've been meaning to make for a while: The default AI does not make mistakes. If your strategy is to hit the enemy as hard as possible and that's what you do, that might not result in the smartest possible move, but it's according to game plan and therefore by definition not a mistake. By contrast, Fred makes all sorts of mistakes. Since he has a variety of possible behaviors, choosing a clearly wrong move does happen. Not much I can do about that, other than slowly chipping away at the causes of the worst of those.

    Anybody still reading this? ;) Sorry, there's no real point here other than that I wanted to write it down to sort out my thoughts a little and for future reference. Comments are welcome, of course, but I'm not expecting any.

My main conclusion on the operations layer work so far: this is quite complex and there is no way that I will get it right on the first attempt. Neither of that is surprising. So instead of working through the entire ops layer task list, I will likely make this a series of smaller releases, each of which concentrating on specific aspects. For example, the next release will likely be mostly about the non-attack actions, and will probably emphasize behavior adjustments against lawful factions.

Or maybe not. We'll see. My thoughts on this change frequently as I work on Fred. Which is part of the fun.
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby The_Gnat » November 27th, 2017, 1:32 am

We all know that playing against the default AI is not how to go about balancing against a human player. No surprise there. What I did not expect was that playing against the default AI is the main reason why Fred is currently so exaggeratedly defensive. However, once I think about it, it makes a lot of sense. Since the default AI just throws everything at you until it is exhausted, sitting back and waiting it out is actually a pretty effective strategy. That's also why the replays you all are providing are so valuable — I simply lack the time to do all the human-player testing I should (and would like to) be doing.


That is an interesting observation. I wonder what tactics the Fred AI would use if fighting itself. Would both sit around until luck proves one AI the winner, or would they make aggressive actions to try to topple the enemy?

I think that the attacks themselves are not the main reason for this defensive play. Sure there's room for improvement there too, but the main problem is not bringing units into positions from where they can attack effectively on the next turn. So that's what I have been concentrating on so far.


Well that somewhat answers the previous question: "Yes" - and therefore shows that it is a good idea to focus on the non-attack actions next.

But I think I need to get it approximately right manually first.


That of course is the struggle when creating AI. I personally can sympathize, though i have never attempted creating an AI of such complexity as required by Wesnoth. But i can say that by the time you are finished with this AI you will most likely be quite an expert at the game ;)

As a side note, once the rough balancing is done, fine-adjusting the weights of these functions is probably a good candidate for applying machine learning and genetic algorithm techniques.


Does Wesnoth's code base currently have any features or modules available that could do this?

We'll have to see what "small number" means in practice. For example, the operations layer now adjusts the base aggression from the config file based on the relative time-of-day adjusted power of the two players. That works well for Northerners against lawful factions, but might not be so great against other chaotic factions. More so, even against different chaotic factions the best overall strategy is not necessarily exactly the same.


True in many situations it is better to fall back (even when the ToD is in your favor) so that you don't loose to many units. I have no knowledge of the wesnoth AI design, but ideally the Tactics Layer (or some form of the tactics system) would allow for conditional changes to the base parameters. For example when many of your units are below half life become less aggressive even at favourable time of day. Is this something that is possible?

Also that brings up a question i was wondering, (This is something fairly basic so i assume it will have already been covered) Does Fred modify his recruits based upon the enemy's faction and recruits?

It is very interesting to consider the possibilities and challenges of AI. :) Recently (for the fun of it) i wrote a AI for a simple Connect Four game. But what i was not able to achieve is making the AI learn from past mistakes. When i have more time i may look into it again and see if it can be achieved (though even in a game as simple as Connect Four having an AI learn as it progresses would take considerable time).
User avatar
The_Gnat
 
Posts: 1384
Joined: October 10th, 2016, 3:06 am

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 27th, 2017, 3:55 pm

The_Gnat wrote:I wonder what tactics the Fred AI would use if fighting itself. Would both sit around until luck proves one AI the winner, or would they make aggressive actions to try to topple the enemy?

Well, it would be a bit of a mix of the two, but the current incarnation would probably mostly sit back and wait for the other side to make a move. The current work will result in a more aggressive overall default behavior, but with the option to choose either via AI aspects.

By the way, Fred vs. Fred testing is on the list of things to be done, but it will take quite a bit of work. Beside the obvious mirroring of all map/side specific entries in the code, there are also more subtle things. For example, Fred takes retaliation on the next enemy turn into account. This includes time-of-day effects, which for Side 1 are the same as for the current AI turn, but for Side 2 might be different. There's also something in the (recruiting, I think) code that makes the AI crash when used for more than one side. All of that can be dealt with, of course, but it takes time and care and I haven't gotten there yet.

The_Gnat wrote:
mattsc wrote:As a side note, once the rough balancing is done, fine-adjusting the weights of these functions is probably a good candidate for applying machine learning and genetic algorithm techniques.

Does Wesnoth's code base currently have any features or modules available that could do this?

No, it does not, but it is not needed either. Since it is possible to run MP games from the command-line in no-GUI mode, external tools can be used for this. Or you can just write your own scripts. In fact, AI-demos still includes machine-learning scripts for recruiting that SeattleDad set up quite some time ago. Those are currently deactivated, but they're still there.

The_Gnat wrote:I have no knowledge of the wesnoth AI design, but ideally the Tactics Layer (or some form of the tactics system) would allow for conditional changes to the base parameters. For example when many of your units are below half life become less aggressive even at favourable time of day. Is this something that is possible?

It's not only possible, but it is something that Fred actively does. And that sort of reasoning is what I am currently trying to improve in the operations layer work. That's exactly what I meant with my comments on the ops layer work consisting of two parts and the "hooks" in the tactics layer. I probably wasn't very clear with that. :) Also just for clarification: that has nothing to do with the default Wesnoth AI. Fred uses the Wesnoth game mechanics, of course, but the code is entirely separate from the Wesnoth AI otherwise.

The_Gnat wrote:Also that brings up a question i was wondering, (This is something fairly basic so i assume it will have already been covered) Does Fred modify his recruits based upon the enemy's faction and recruits?

Yes, he does, but the recruiting still has lots of room for improvement (and yes, you're right, others have pointed that out before too). Currently Fred uses a only very slightly modified version of the recruiting code of the Experimental AI from mainline (which is completely separate from the default AI recruiting code).

The_Gnat wrote:It is very interesting to consider the possibilities and challenges of AI. :) Recently (for the fun of it) i wrote a AI for a simple Connect Four game. But what i was not able to achieve is making the AI learn from past mistakes. When i have more time i may look into it again and see if it can be achieved (though even in a game as simple as Connect Four having an AI learn as it progresses would take considerable time).

Indeed. In my experience, at least for Wesnoth, you really need to set things up to work pretty well manually first. The learning part might then be useful for fine-tuning, but I am not there yet with Fred. Have fun with your own AI work!
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby Cold Steel » November 28th, 2017, 6:59 pm

mattsc wrote:That's also why the replays you all are providing are so valuable — I simply lack the time to do all the human-player testing I should (and would like to) be doing.


I have been meaning to ask, what is the optimum number of replays you would like to see for each release, in a perfect world? I think we have been averaging about a dozen per, which can be increased, but at want point would you be swamped with replays? Also how many replays would you want from a particular player, maximum?

mattsc wrote:the main problem is not bringing units into positions from where they can attack effectively on the next turn. So that's what I have been concentrating on so far.


Maybe a generic aggression modifier or weight could be set both per opposing faction && per each of the six specific times of day.

So if need be, you could weight for dusk, +20% aggression when against loyalists but +25% against drakes (to compensate for their retreating mobility as an example). You could also make for most lawful or neutral faction match ups, aggression +50% during first watch and then fall to 40% on second watch, such that if not enough damage is done on the first turn of night (which lowers the threat level of the opponent's forces) he can begin getting just a bit more defensive to keep from getting encircled. It would also cover the versus-chaotic-factions-matches as well, of course.

The downside to this approach is the number of weights grows as Fred becomes more generic. Right now, there are 6 times of day by 6 opposing factions, so 36 weights. But when Fred can play other factions too, that balloons to 216 weights. When he can play more than one side, you may or may not need twice as many weights to compensate for added consideration of who moves first after each ToD change. And when he can play 1vs1vs1 or 2vs2 maps, who knows.

mattsc wrote:As a side note, once the rough balancing is done, fine-adjusting the weights of these functions is probably a good candidate for applying machine learning and genetic algorithm techniques. But I think I need to get it approximately right manually first.


Do you mean machine learning that pulls data from Fred versus Human matches or Fred versus Fred in no GUI mode (times tens or hundreds of thousands of matches)?

mattsc wrote:Currently Fred uses a only very slightly modified version of the recruiting code of the Experimental AI from mainline (which is completely separate from the default AI recruiting code).


I have been wondering, what happens when you plug in the recruiting code from the default "Strong AI" into the Experimental AI (Ron) or Fred? I imagine they fair less well against a default AI opponent but do they do better against humans?
Cold Steel
 
Posts: 261
Joined: January 6th, 2008, 3:32 am

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby singalen » November 28th, 2017, 7:03 pm

Oh, so you need replays in numbers? Will provide more.

BTW on the latest dev build "AI Demos" gives an error about undefined macro WESNOTH_VERSION - I wonder if it's a problem in Wesnoth or in AI Demos.
I got it on iOS.
User avatar
singalen
iOS Port Maintainer
 
Posts: 178
Joined: January 3rd, 2007, 10:18 am
Location: bay

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 28th, 2017, 7:11 pm

Hi All -- thanks a lot for all the comments. I'll reply in more detail sometime in the next 48 hours, for now just this:

singalen wrote:Oh, so you need replays in numbers? Will provide more.

Actually, I have what I need right now. I still haven't made full use of what you all have posted and the next version is likely going to play quite differently (actually, the current development version already does), so replays using 0.14.8 will only be partly applicable except for setting up testcases (of which I have plenty from the previous round).

Sorry, I have to run off right now, just wanted to say this quickly before you go through too much trouble. I really appreciate all the help though!!
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby Shiki » November 28th, 2017, 8:10 pm

singalen wrote:BTW on the latest dev build "AI Demos" gives an error about undefined macro WESNOTH_VERSION - I wonder if it's a problem in Wesnoth or in AI Demos.

It's a problem in wesnoth. You can get it if you have this macro in your code, have another error, fix it and restart wesnoth. Shadowm mentioned it's because something corrupts the cache, introduced in or after 1.10. Not being worked on.
Maintainer of Era of Myths.
Shiki
 
Posts: 146
Joined: July 13th, 2015, 9:53 pm
Location: Germany

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby singalen » November 28th, 2017, 9:34 pm

Shiki wrote:It's a problem in wesnoth. You can get it if you have this macro in your code, have another error, fix it and restart wesnoth. Shadowm mentioned it's because something corrupts the cache, introduced in or after 1.10. Not being worked on.

Oh, thanks a lot! It might be a pointer for my on work.
User avatar
singalen
iOS Port Maintainer
 
Posts: 178
Joined: January 3rd, 2007, 10:18 am
Location: bay

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 29th, 2017, 4:06 pm

Cold Steel wrote:I have been meaning to ask, what is the optimum number of replays you would like to see for each release, in a perfect world? I think we have been averaging about a dozen per, which can be increased, but at want point would you be swamped with replays? Also how many replays would you want from a particular player, maximum?

Hmm, that's difficult to answer. I should probably give some background on what I actually do with the replays first. I watch them of course, just to get an impression of how Fred is doing. Then I watch them again in more detail, trying to pay attention to where there were the biggest problems/mistakes. I then save the game from the replay at those turns and go into manual testing mode. AI-demos has debug functionality built in with which I can take a save and evaluate and/or execute individual moves manually while displaying a bunch of debug information. That lets me figure out why Fred is doing whatever he's doing, and then work on trying to fix it, re-evaluating/executing the same moves over and over again. Then I can modify the situation on the map in debug mode or find similar test cases in the other replays (or from my archive of all test cases), to make sure I did not just optimize for that one specific situation, but that it is somewhat robust to changes. Anyway ...

So, given that so far we are still mostly working on getting the broad strokes roughly right, one or two replays per faction is probably about what I need. Meaning, what you've been doing is pretty much perfect. Once we get into fine tuning the behavior, it might be more useful to get more replays of the same type (e.g. a bunch of replays against Drakes, rather than one per faction), but I don't think we're there yet.

I don't think there's a limit for what is useful from anyone individually. Having a variety of playing styles from different players available is good, but seeing how consistent Fred's behavior is when pitched multiple times against the same player is very useful too. Mostly, I don't want anybody to feel obliged to have to provide anything. This should be fun, not a burden (which is why I sometimes take week- or month-long breaks myself). :)

Cold Steel wrote:Maybe a generic aggression modifier or weight could be set both per opposing faction && per each of the six specific times of day.
[...]
The downside to this approach is the number of weights grows as Fred becomes more generic. Right now, there are 6 times of day by 6 opposing factions, so 36 weights. But when Fred can play other factions too, that balloons to 216 weights. When he can play more than one side, you may or may not need twice as many weights to compensate for added consideration of who moves first after each ToD change. And when he can play 1vs1vs1 or 2vs2 maps, who knows.

Yeah ... Those are all good suggestions and roughly what I am trying to do. But I am also trying to avoid what you say in your last paragraph. So currently the code is trying to figure this automatically, at least approximately. The AI gets more aggressive when the power shifts due to time of day. It also looks one turn ahead and does a partial adjustment for the power ratio on the next turn. This is done based on current units on the map, not theoretical power of the faction as a whole. Thus, for example, Fred against lawful factions is a bit more aggressive at dusk than at dawn, and a bit more aggressive during first watch than during second (unless he managed to gain more of an advantage during first watch, just as you said). But of course, that will likely not fully account for the differences between Loyalists and Drakes, or Northerners and Undead.

In any case, for simplicity I'd prefer to keep it this way for the "general" version of Fred. But then there will also be lots of configuration parameters to adjust this for a given map or scenario, and it is trivial to make those dependent on faction or time of day. I'd rather have that as an option though, not the default.

Apparently I'm on a rambling trip again. ;) Hope this makes at least some sense.

Cold Steel wrote:Do you mean machine learning that pulls data from Fred versus Human matches or Fred versus Fred in no GUI mode (times tens or hundreds of thousands of matches)?

I mean Fred vs Fred in automated batch mode. I think you need a huge number of games for this and it won't be practical to get those from games vs humans.

Cold Steel wrote:I have been wondering, what happens when you plug in the recruiting code from the default "Strong AI" into the Experimental AI (Ron) or Fred? I imagine they fair less well against a default AI opponent but do they do better against humans?

Well, the strong AI really just uses the default AI's recruiting code with slightly modified parameters. Ron/ExpAI/Fred's recruiting code is vastly superior to that. If you do nothing but using the ExpAI recruiting code, it beats the default/strong AI somewhere around 2/3 of the time (I forgot the exact numbers, we did those tests many years ago). If you use all of the Ron/ExpAI candidate actions, the win rate goes up to >80%, and of course you know how Fred does.

Edit: I just noticed that I did not answer your last question: No, the percentages change against human players, but the trends are the same. As in, both ExpAI and Fred, in spite of their current problems, definitely do better against humans than any version of the default AI (incl. the "strong" AI). That includes the recruiting (at least on average, there might be individual cases where that is not the case).
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby Poison » November 29th, 2017, 6:03 pm

mattsc wrote:The AI gets more aggressive when the power shifts due to time of day. It also looks one turn ahead and does a partial adjustment for the power ratio on the next turn. This is done based on current units on the map, not theoretical power of the faction as a whole. Thus, for example, Fred against lawful factions is a bit more aggressive at dusk than at dawn, and a bit more aggressive during first watch than during second (unless he managed to gain more of an advantage during first watch, just as you said). But of course, that will likely not fully account for the differences between Loyalists and Drakes, or Northerners and Undead.

In any case, for simplicity I'd prefer to keep it this way for the "general" version of Fred. But then there will also be lots of configuration parameters to adjust this for a given map or scenario, and it is trivial to make those dependent on faction or time of day. I'd rather have that as an option though, not the default.

Apparently I'm on a rambling trip again. ;) Hope this makes at least some sense.


Wow this is really nice, it sounds kinda like a human player. In the version I've played it didn't seem to be like that (as I have pointed out), I'll definitely try the new one soon. Everything makes perfect sense, don't worry.

OT: Damn your Galuldur campaign is difficult, hard is definitely nightmare. I've managed to get past scenario 3 (-1%,-2%) eventually (partly using your advice and creating a marksman - I had replayed the first scenarios again- I don't like marksmen in general, I prefer rangers, they seem more survivable but he was really useful) only to find myself even more frustrated with the 4th one :( I do have an idea on how to beat it, though. Despite the fact that the campaign is (barely) beatable on hard, I think a little better gold management would be useful for some scenarios to give the player some breathing space and cope for bad luck. If you ever find yourself to agree and have some spare time invest a little on improving that too. Your effort is truly appreciated.
User avatar
Poison
 
Posts: 87
Joined: August 13th, 2017, 4:54 pm

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 29th, 2017, 6:24 pm

Poison wrote:
mattsc wrote:The AI gets more aggressive when the power shifts due to time of day. It also looks one turn ahead and does a partial adjustment for the power ratio on the next turn. This is done based on current units on the map, not theoretical power of the faction as a whole. Thus, for example, Fred against lawful factions is a bit more aggressive at dusk than at dawn, and a bit more aggressive during first watch than during second (unless he managed to gain more of an advantage during first watch, just as you said). But of course, that will likely not fully account for the differences between Loyalists and Drakes, or Northerners and Undead.


Wow this is really nice, it sounds kinda like a human player. In the version I've played it wasn't like that (as I have pointed out), I'll definitely try the new one soon.

Yeah, this is new since the last release and it still is a bit "rough around the edges" with some undesirable side effects that I would like to iron out before releasing the next version. Either way, it will definitely be in v0.14.9.

Poison wrote:OT: Damn your Galuldur campaign is difficult, hard is definitely nightmare. I've managed to get past scenario 3 (-1%,-2%) eventually (partly using your advice and creating a marksman - I had replayed the first scenarios again- I don't like marksmen in general, I prefer rangers, they seem more survivable but he was really useful) only to find myself even more frustrated with the 4th one :( I do have an idea on how to beat it, though. Despite the fact that the campaign is (barely) beatable on hard, I think a little better gold management would be useful for some scenarios to give the player some breathing space and cope for bad luck. If you ever find yourself to agree and have some time invest a little on improving that too. Your effort is truly appreciated.

Thanks for those nice and useful comments too. I'm glad you made it past S3. Unfortunately I haven't had time to look into it myself yet. Fred definitely has priority over Galuldur at the moment and I've also chased down a couple mainline bugs over the last few days. Well, and then there's real life ... Anyways, I do believe that "hard" should actually mean hard, but Galuldur is not meant as an expert-level campaign, so I'll make sure to look into it. I'm not opposed to making it easier, I just don't want to change the character of the scenarios.
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby Cold Steel » November 30th, 2017, 3:51 pm

mattsc wrote:So, given that so far we are still mostly working on getting the broad strokes roughly right, one or two replays per faction is probably about what I need. Meaning, what you've been doing is pretty much perfect. Once we get into fine tuning the behavior, it might be more useful to get more replays of the same type (e.g. a bunch of replays against Drakes, rather than one per faction), but I don't think we're there yet.


Works for me. When that time comes I could also start hosting some games against Fred in Wesnoth multiplayer, to bring in more tester variety.

I did something similar with Ron a while back for less productive ends. I would join Auction-X games and suggest the host switch the center Default AI to Experimental AI. Everyone would die. :D But eventually you started to see most Auction-X games, hosted by others besides myself, feature Experimental AI for about a month afterwards. People were so challenged and impressed by how much more difficult it was that they had to keep playing it until they could find ways to overcome it. Quite a lot of fun.

mattsc wrote:In any case, for simplicity I'd prefer to keep it this way for the "general" version of Fred. But then there will also be lots of configuration parameters to adjust this for a given map or scenario, and it is trivial to make those dependent on faction or time of day. I'd rather have that as an option though, not the default.


Makes sense.
Cold Steel
 
Posts: 261
Joined: January 6th, 2008, 3:32 am

Re: Help with testing Fred - Freelands MP Custom AI v0.14.8

Postby mattsc » November 30th, 2017, 11:53 pm

Cold Steel wrote:Works for me. When that time comes I could also start hosting some games against Fred in Wesnoth multiplayer, to bring in more tester variety.

I did something similar with Ron a while back for less productive ends. I would join Auction-X games and suggest the host switch the center Default AI to Experimental AI. Everyone would die. :D But eventually you started to see most Auction-X games, hosted by others besides myself, feature Experimental AI for about a month afterwards. People were so challenged and impressed by how much more difficult it was that they had to keep playing it until they could find ways to overcome it. Quite a lot of fun.

Huh, that's cool! Thanks for letting me know (and doing so in the first place), good to hear that that worked and that others found this interesting/challenging. It would be nice to get Fred some more exposure at some point, but my guess is that we're still at least 3 or 4 releases away from something similar making sense.
mattsc
 
Posts: 991
Joined: October 13th, 2010, 6:14 pm
Location: Wandering, mostly aimlessly

PreviousNext

Return to Multiplayer Development

Who is online

Users browsing this forum: Mint, Yandex [Bot] and 1 guest