I've just watched the siege AI presentation at EGX rezzed.
https://www.youtube.com/watch?v=fj05Ll1Mark - Video
http://gdcvault.com/play/1023038/Hav...ing-the-Castle - presentation slides
An excellent presentation that went into a good level of depth, for anyone who is interested in how the game itself works this is a must watch video.
It also showed off some good new detail on certain units e.g. chaos spawn and cairn wraiths
Well done CA on their transparency and communication.
However, there were issues in the way that the battle AI is intended to work, which this presentation highlighted.
Because this was a good presentation which exposed the inner workings of CAs product, any criticism should be constructive, so that CA don't feel that releasing this video was a mistake.
Right, now that's been said onto the criticism.
Design philosophy
There are five warhammer siege goals as stated by CA:
- Fast high intensity battles
- single attack direction
- focus the battle on the city walls
- quick resolution once past the walls
- press the defender quickly and broadly
CA have said a few times now that their intention is to create "fast battles with high intensity", this was repeated today. I suspect this is a major factor in the fast battles in previous games which receive a lot of negativity (including from myself). This isn't a fault with the finished product, because the finished product works as intended.
Depending on your point of view it is a problem/feature of the design philosophy. What this means is that nothing is broken/bugged and this will be present in the finished product.
Its also worth noting that 3 of the 5 goals there mention battle speed, this is obviously a major concern of CAs. I personally enjoyed the slower battles from previous total wars so it would be interesting to know what are the thoughts behind CAs slow battle concerns.
CA have said that "the walls are key" and that once you are past the walls it should be a quick finish, again this may result in less street battles, which could be an issue if you enjoy a battle where you have to fall back to a secondary defense point, but from the videos we have seen, the street battles don't seem to be reduced too much.
I suspect they have positioned the victory point closer to the walls so that the attacker doesn't take as long to reach the defenders at the choke points.
This could help the AI's separate attacks to happen at the same time, instead of a piecemeal stream of units.
Personally I think a closer victory point could be beneficial to the game, but it remains to be seen.
"quick to battle", this philosophy is probably behind the defensive towers being able to attack the entire deployment zone, which pushes the attacker into assaulting quickly before loosing too many units to ranged fire.
Personally I think this is a mistake, particularly if you like artillery heavy armies such as the empire and dwarfs.
and it is probably behind "pocket ladders" which means every unit has access to a method of climbing the walls without having to slowly walk with siege equipment.
In a siege the attacking force has absolute control over when the attack takes place, unless the defender chooses to sally forth.
By pushing "quick to battle" onto the player, CA are pushing certain tactics to be more viable than others, instead of allowing the player to make their own choices about how they want to play the game. I think this will feel more frustrating for the player rather than challenging.
Personally I think it is flawed because the attacking player doesn't have control over the attack. The defender has control over the attack by choosing to occupy certain sections of wall or not. (occupying certain sections of wall activates the defensives towers).
Overall strategy implementation
CA have shown us the architecture of how the battle AI and siege AI looks at the overall strategy.
The "grand tactical analyzer" works for the battle AI and looks at the overall battle and determines the objectives that each detachment (group) of units will perform.
The intended purpose of each subsection of the grand tactical analyzer seems rational, it seems like CA have a decent architecture in place to provide a good AI.
Where the AI fails is probably due to exceptions which the logic doesn't consider, design philosophy and the inability to call off an attack.
Exceptions which the logic doesn't consider are always going to be present in games until an AI that actually learns from previous mistakes has been created. Unfortunately this is a case of repeated testing, feedback and patching. These are typically the "gamey exploits" that get used by the human player.
The design philosophy covers heavily scripted events such as quest battles, e.g. if CA script the AI to be stupid then it will be stupid.
and such features as the "fast high intensity battles" so If CA want the melee combat to start quickly, then the AI will try and engage melee combat early, regardless of whether it would be beneficial for the AI to hold back with a ranged advantage (see azhags quest battle).
It was mentioned that "the AI wont ever withdraw", which can be seen in previous games for example in shogun 2 you would successfully defend against a siege and still have most of your army left, but the AI would still continue the attack even when it had only the general unit left against a near full army.
The AI not being able to withdraw is a fundamental flaw in the overall strategy.
Fleeing from battle is not withdrawing, withdrawing would be a controlled walk to the edge of the map (so that they could defend themselves) then fleeing, to preserve as much of their army as possible.
When the AI can be exploited/gamed due to unforeseen exceptions, but must continue the attack because it cannot withdraw, then the AI will always perform badly, these are fundamental failures of TW AI that have been present in many games.
e.g. rome 2. you would set up a unit of pikemen in a chokepoint, which was an overly powerful defense, and the AI would not be able to withdraw after suffering enough losses to realise that its tactic is not working.
Gamey tactics will still be very much possible in TW:W, unless the AI is given the ability to; withdraw, call off an attack or learn how well that tactic is working.
Also, if the AI was attacking a city it could not begin the assault, suffer enough losses to realise the assault was not going to work and then withdraw and maintain the siege on the campaign map. The AI has to fully commit with the attack, regardless of circumstance.
Specific tactics
it was stated that the grand tactical analyzer is not present in sieges due to the tactics being the same during siege attack regardless.
e.g. take the walls, then take the victory point.
CA showed the layout of options that were available to the AI and "holding back and bombarding" will not be possible for the AI as it is not an option and because it has no way of deciding to hold back or advance forward because it doesn't have a grand tactical analyzer.
The AI will attempt multiple simultaneous attacks, one of which is bombarding the walls to break the walls and create entry points.
However, the AI could have a huge ranged advantage, which could cause your defending melee heavy army major losses, but they will still try and melee the walls from the beginning, regardless. They have no way of considering an initial melee assault a bad idea, due to no grand tactical analyzer.
This will be terrible for AI armies which have strong artillery and weaker melee infantry. The melee infantry will have begun their assault and may be thrown away by the time the artillery has done damage to the walls and units. In the demonstration battle they showed the artillery destroyed the wall after the melee infantry had reached the walls and engaged in combat.
When this is paired with the inability to withdraw then the siege AI it is effectively just throwing its units at you regardless of circumstance.
The one redeeming feature is the implementation of a "reserve pool", this allows the AI to withdraw units out of an assault (based on circumstance), back to the reserve pool and back into a different assault.
However, this will not stop the AI from throwing most of its melee infantry at you from the beginning regardless.
It also does not give the artillery time to be useful.
Siege battles will feel very similar due to the AI not being able to consider alternative tactics, particularly with similar layouts between cities.
It also will be fairly easy for the human to recognise the patterns and create gamey tactics to counter those patterns.
I also did not see any option for the artillery to destroy the defensive towers, it seems like it will just attack the walls directly infront of it and then units in the city.
The only options I saw for the artillery (and the rest of the units for that fact) were:
Assault walls, assault gates, breach walls.
Hopefully this doesn't result in heavy artillery losses and the AI not doing anything about it due to the defensive towers ranged attacks, once the rest of the units are past the walls. if you keep a unit on that section of wall.
CA showed how they set the perimeter of the city so that the AI knows where the walls are. This is so that the defending AI knows the perimeter to stay within. CA did say that the image was a representative and not an actual image, but it doesn't seem to work very well for the gates.
The 2D perimeter sits where the outside of the walls are, but it doesn't take into account the gate being slightly in from the walls.
The perimeter sits outside the gate, which allows the AI units to sit outside the gate, this is possibly responsible for the glitches in previous games where half of an AI unit would stick through the gate. either way I think it might have been sorted in later games, but if not, I hope this has been addressed so that units stay within the perimeters when they are meant to.
With that in mind, I don't believe the defensive siege AI has the option to send out cavalry or fliers to take out enemy artillery, but the defensive siege AI wasn't really discussed as much, so It could be possible, either way I hope it is. Because it is a viable tactic, especially with fliers.
In the event of a wall being knocked down, the AI spots this and send units from the reserve pool into this breach, there was no mention of considering which units were on the other side of the breach, e.g. a unit of pikemen in phalanx.
I hope they consider which units are defending the breach, so that they don't throw units away e.g. charging cavalry into a phalanx.
Reserves
This being the new feature of TW battle AI deserves its own special section. Personally I think it is a fantastic feature that allows the AI to re-analyze decisions that have already been made. The intention is excellent.
CA said, that if a tactic fails, units are sent to the reserve pool. Although it was ambiguous what causes a tactic to fail, I believe CA meant, things such as, if the siege tower is destroyed then the 2nd and 3rd unit that would have assaulted the walls via that tower would return to the reserve pool, to be allocated to a new/existing tactic.
I would like to see some clarification about what causes a tactic to fail though.
The main issue that I have with the reserves pool, is that CA said that units have to walk to a specific area (reserve pool area) before being re-assigned to a new tactic.
This means that a unit at the walls that has lost its siege tower would have to walk all the way back to the deployment zone before walking all the way back to the walls as part of a new tactic.
This is easily exploitable, e.g. hold fire on the nearly dead tower, until the tower is nearly at the walls, then destroy it so that the units spend longer walking back, which exposes them to more ranged fire.
Also they will tire more through walking more, and the reserves which are meant to reinforce certain tactics will not be able to reinforce those tactics when they are needed most, they will have to walk back and forth before reinforcing.
This reserves pool walking back and forth was very evident during the recent chaos/empire field battle. There was a heavy focus on bringing back a unit to the reserves pool as soon as the previous reserves allocation has been performed, there was little consideration that all of the units are currently in the right place so let them stay in place, until they can be brought back to the reserves pool.
I've provided evidence and a breakdown in this post.
http://www.twcenter.net/forums/showt...1#post14927324
Summary
There were plenty of decent features mentioned here, but I haven't touched on them because they seem to work. I've only wrote about the features that need further work/consideration.
As always during the recent few games, CAs design philosophy is called into question, particularly with regards to fast battles etc.
This is a case of personal preference, but I would like to see a better effort on CAs part to accommodate for all tastes, e.g. battle speed settings.
The individual tactics of the battle AI does seem to be improved over the previous games AIs, which is particularly good seeing as we are at pre-launch. However their are plenty of exceptions that need to be accounted for.
The overall tactics of the battle and siege AI seems to have major issues that will result in throwing away units, which we have seen in a few videos so far and a lot in previous games.
That being said, they do look like they have the architecture in place to be able to fix the existing AI, rather than creating a new architecture which would have its own problems.
It sounds like the AI needs a few more tactics to choose from, e.g. withdrawing from battle.
CA need to make their list of exceptions per tactic as complete as possible, which means plenty of testing, plenty of feedback and plenty of patches.
Seeing as we are getting 3 games ,which all have to work together, CA should have been able to spend more time testing this AI at the end of the trilogy than any previous games, which hopefully should result in many improvements, unless their design philosophy blocks certain improvements.
The reserves pool is a welcome addition to the game and given enough testing and patches should become a good feature.
Conclusion
CA have the basis to create a good AI, but need to remove the design philosophy influences that hold the AI back. Testing and patches can fix the bugs/exceptions with the AI, but they need to be allowed to fix those errors in the first place, which seems to be what is holding the AI back.