What need an AI do?

[MarkShot]

MY QUESTION PART I:

In the thread entitled "PC Civil War Games", there has been much discussion of about AIs.

A lot of people who play mainly SP (single player) have indicated that a strong AI is a key component of a worthwhile game.

Well, I wanted to learn more about people's perception of what a good AI is. Of course, simplistically, we can just say that a good AI is one that beats human players without cheating. (Although I would might argue that cheating is okay as long as it is not extremely overt and leads to an exciting game ...).

So, what must an AI do to be considered good? I am looking for detailed answers that identify fairly specific military strategies that lead you as the human player to feel that the game is immersive. That the AI is capable of formulating a plan and executing it. So, more than simply massing and hitting a weak spot on the line. Here let me throw out some examples:

AI knows how to:
- Use terrain (will vacate poor terrain; will defend on terrain with bonuses; will avoid attacking good ground and instead look to bypass).
- Feints or masses in one location while planning the real attack elsewhere.
- Knows how to withdraw to pull an enemy forward streching supply lines and then counter attacks looking to encircle/cutoff an army.
- Use seasons; will attack in the Spring and use the Winter to consolidate and secure its gains.
- Will set up a layered defense such that the enemy is gradually attrited as it advances.
- Will use blocking and delaying forces to slow an attack so that the main defensive line can prepare its positions.
- Will maintain a reserve for various uses like a counter attack when the enemy is spent.
- Will use pickets to gather intelligence about the enemy.
- Will use screen forces to deny the enemy intelligence about actions and movements deep in the rear.
- Will give priority to the transportation network such it will attempt to seize key junctions denying the enemy ability to move troop and reinforce a battle.
- Will peform a hasty attack if it perceives that the enemy is disorganized; will perform a careful cohesive attack against a prepared defense.
- Will use terrain as part of a defensive plan like set up a defense on the far side of a major river; give priority to holding bridges.
- Will use exploit a break in the lines with fast forces such as cavarly to cause havoc in the rear.
- Will give priority to political objectives that might break the will to fight.
- Will attempt to flank a line a roll it up.
- Will draw the enemy into kill areas made up of difficult terrain, prepared defenses, and artillery barrages.
- Will trade terrain for time if the future offers recruitment/trained troop/foreign reserves.
- Will preserve its forces and avoid attrition situations if manpower levels are low.
- Will force high attrition combat if the enemy has less manpower.

Well, the above are just some examples. The point is that the items above are what you would expect of a human commander.

So, when you say about an operational/strategic game "good AI!", what does the AI need to do for you to say that?

Thanks.

---

MY QUESTION PART II:

Now that you have a sense the level of play/command that qualifies a worthy opponent in my mind ... For those of you who feel that the AGE engine has a good AI, what speficially does the engine do that makes you say this? Where is the engine strong in its ability to analyze the situation and attack or defend? Where is the engine weak? What situations arise where you find yourself smirking saying "dumb AI!"? (sorry Pocus)

What situations occur where you manage to kick the AI's butt repeatedly? What is the AI failing to see or do in order to avoid repeated embarassments?

Thanks.

---

Thanks for your time to take my questions seriously.

[Jbeoddy]

Now this is just off the top of my head, but here goes.
First, Athena is probably as good an AI as I've seen in a war game.
That having been said:
A 'superior' AI might demonstrate the ability to conceptualize and carry out long term 'strategic' plans. (In my humble opinion, one of the great flaws of the Southern Command was that their 'long-term' strategy appears to have been 'stand on the defensive and beat the tar out of any Federal army that came south'. Whereas the Union decided on and carried out the 'Anaconda Plan almost from the beginning of the war, even in the face of reversals, and outright catastrophe, normally brought on by the seeming timidity of the Union Generals. All of the Confederate 'incursions' were typically for short-term limited objectives, the invasion of Maryland in 1862 was aimed at bringing Maryland into the Confederacy on the theory that they were a Southern sympathetic state suppressed by the Union. The invasion of Pennsylvania in 1863 was aimed at drawing off strength from the western campaign in an effort to relieve Vicksburg. Tactically brilliant, strategically meaningless.)

A 'superior' AI might have the ability to conceive strategies or tactics that maximize the strengths of the position and minimize their weaknesses. For example, the Union strategy in the actual war emphasized their naval superiority. It minimized the usage of cavalry til late in the war. Thus, their strength, in the form of a strong nucleus of a navy was matched against a confederate weakness, while a weakness in their cavalry, where, early in the war they were overmatched both in personnel and leadership was minimized.

A 'superior' AI might show an ability to perform better 'risk analysis', such as opting not to pursue with troops that had suffered severe attrition, when the situation into which it was pursuing was unknown. (This is a flaw I think Athena is guilty of sometimes. If I set a 'weak' unit, say a division, in a relatively exposed area and Athena defeats it, she will almost invariably pursue. Typically running right into the corps that that division just rejoined, which is typically flanked by two more with the army in reserve. Which often costs her an entire corps, if there's a cavalry division prepared to move in behind her. Or even worse, an entire cavalry corps. I've caught her with that three or four times.)

A 'superior' AI might show a better abililty gauge strategically important positions. Examples of this in Athena's case abound throughout the forum, as people talk about 'gamey' strategies pursued against Athena, mostly involving drawing her out with a diversionary attack on a strategically meaningless (relatively) objective and leaving Washington or Richmond wide open. Historically, one of the problems with Lee's two invasions was that they were obviously 'demonstrations only' as no serious observer would have believed any Confederate attack on Washington could succeed. After the debacle at first Bull Run, the Union defense of Washington was probably never less than a 100,000 men in multi-layered entrenchments, with the heaviest guns they could mount, inside brand spanking new 'modern' fortifications. Any assault on Washington which was not preceded by a campaign inflicting incredible attritional casualties on a Union Army that was 'in being' and stood some chance of surviving if reinforced speedily just wasn't going to succeed. However, Athena (and it might also be said, almost any AI in any game you care to name) can be convinced to draw troops out the Washington defenses in any number of cases. (For the historical purists, try creating the massive armies that Athena typically throws at the Confederacy and STILL keep a 100,000 man army with massive Artillery in Washington. This may be part of the answer to the questions about 'ahistorically large' armies by both sides. In all of the AAR's I've seen very little attention is paid to 'rear area security' and almost all of the action is crustal defense, rather than defense in depth, when in reality, both sides, but the Union moreso, held massive numbers of men in reserve, defending strategically important objectives.)

A 'superior' AI might learn from it's mistakes. Reading through the forum, it's apparent that some players have found 'gamey' strategies for defeating the AI, particularly when they are playing the south, due to some pecularities in the AI's assessment of threats. The one I'm thinking of is the one where they take advantage of the one region adjacent to Washington with no river barrier and no rail, which apparently reduces it's importance in Athena's decision tree. A 'superior' AI might fall for this once, but learn and not repeat the mistake. (One DISADVANTAGE of this might be an AI that learns might become 'unbeatable' over time. In which case, I'd suggest pulling the plug before we're all living in "The Forbin Project". )

Any improvement in these would represent a giant leap forward. Do I think these could be implemented? Not so much. Of the bunch, perhaps the easiest to implement might be the one about improving defense of strategically 'important' places. This would probably revolve around increasing some 'value' already in the engine. Or a multiplier or somesuch. (Not knowing the algorithm used to determine 'strategic importance handicaps me in knowing how 'easy' it might be. I'm just blue-skying on this one.) The others, well, if y'all were good enough to program that in, we'd be reading articles by you in all the technical journals after you'd collected your Nobel Prizes for creating a true Artificial Intelligence. (And probably reading articles published by Athena as well. Right after she cured cancer, solved the problems in the Middle East, and came up with a faster-than-light drive. ) The point being (now that I've got my tongue out of my cheek) is that I suspect it's a lot harder than it might seem.

Now as to why I think Athena is a "superior AI", at least in comparison to most of the ones I've seen:

Most of the time Athena at least seems to be TRYING to think strategically. She can sometimes be 'diverted', the strategy may not be the best one possible, it may not take advantage of her sides strengths. But at least she's trying. Most of the AI's I've seen have no concept of the difference between 'tactical' success and 'strategic' success.

Athena's behaviour can be 'optimized'. If I think that reducing the fog of war for her causes her to be too aggressive in 'pursuit', for example, I can mitigate that to some extent in the next game by adjusting a setting and reducing her ability to 'penetrate' the fog. I can force her to live with the 'historical' traits of her generals, more or less, by adjusting the activation bonus. Generally speaking, there is more ability to 'fine tune' the AI than I've seen in any other game engine. If there's any 'drawback' to this, it's that you have to play a few games in order to get a 'feel' for what a particular adjustment will do, AND how it may interact with the other optimizations.

The AI WILL react differently in different situations. What I mean here is that if my strategy changes, so does Athena's. I normally play the South. And I generally play a fairly 'defensive' game the first year, while I organize my forces, trusting that even if Athena manages to get McDowell or McClellan to move, they'll just be coming south without trained troops, with low cohesion, and inadequate coordination. All of that being 'inherent' for both sides early in the war. And the defensive being stronger than the offense. Even so, there are 'variations' within that. For example, I might build up a strong cavalry division under McCulloch in Texas and embark on the 'grand tour' of the western territories. If I start by taking Tucson, (almost a must for this scenario to have any chance of success), then Athena WILL react to that. (If I start by going for Denver, she reacts, but late, and not the same way.) Same if I force action in the Tennesse/Kentucky theatre. If I don't force action there, Athena has a pretty standard sequence of actions, but they vary sufficiently that they aren't 'predictable' in the sense that she may always attack Lexington, but not necessarily always in the same force, and what she does thereafter is subject to change, based on what, I'm unable to determine.

This has been an interesting exercise, but when it comes right down to it, my feeling is that while it would probably be possible to 'tweak' the AI in a couple of places, with varying degrees of difficulty, I doubt that it could be improved a whole lot. The suggestions I had for a 'superior AI', if they could be implemented, would likely make the game engine far more valuable, as I could get you a REALLY good contract with the Pentagon. I suspect that Athena's ability to think 'strategically' is probably as good as it gets. Simply because the number of variables is way too high for straight computation to achieve an answer in a reasonable time. (If you were to decide to write one for a heuristic 'neural network' type computer, it would probably be a LOT easier. Of course, I suspect that the potential market might be a bit limited. Actually a LOT limited. Same with 'learning from mistakes'.

It might be possible to improve her 'risk assessment' but adjustment of the 'fog of war' seems to work almost as well, once you figure out what that affects and adjust those parameters as well. (I find, no fog of war reduction, normal aggressiveness, some activation bonus, works pretty well, although it can be interesting to change one or more of these just to see how it works out, even if not 'optimal'.)

Improving the 'survival instinct' in defending important objectives might be doable, but if were me coding it, I'd want to do some extensive testing to see what kinds of alterations this made in her behaviour. (Might cause her to be so cautious that it wouldn't MATTER if McDowell ever activated, she wouldn't have given him a large enough field force to do anything with. )

There's my two cents worth. (Actually more like $1.95. Which, if converted to Euros would probably look like two cents again. )

[GShock]

Considering the vast superiority in number and resources of the Union, the fact the AI attacks and loses, wasting men, NM, resources and giving CSA player bonuses afterwards, i think the AI should be more aggressive with inferior forces and more cautious with stronger ones. So, essentially, the AI makes mistakes when assessing the player's strenghts, attacking when it shouldn't and not attacking when it should.

The AI is heavily handicapped by the activation rules...unfortunately i still believe the best solution against a player who detaches an inactive general and proceeds with an independent stack, is to lock automatically stacks with an inactive general and i mean locking the general to the stack so it can't be unattached. Seems the only way and i'm sure as handicap on both sides this gives the AI a strong needed help.

With that superiority in mind, for the Union, i am amazed there's still no ability for the AI to produce amphibious assaults. This is by far what i consider the top priority in bettering the AI. Maybe it can be achieved via events, maybe with a "transport box" like with the blockade/shipping boxes but something must be done about this problem.

So to recap...balance the AI attacking aggressiveness with the rough calculations the AI can do about win/lose ratio and increase the accuracy of this calculations so the AI attacks when it's got reasonable chances to win.
Add a solution to the total lack of Union AI amphibious assaults.

Of the 2 i think the latter is more important bc with an amphibious capable AI, the CSA can't defend other places as it does right now. So everything would improve.

[Pocus]

Interesting thread. From Mark's list, I would say the AI does 1/3 of it seriously and is in the beginning of doing perhaps another 1/3 within one year

I do think the AI in the engine can still be improved. It has just been improved some weeks ago by reworking, cleaning and double-checking many of the low level code it used, and progress can still be made. There are now some new tools available, which are not yet used to their full potential, in the AI code. This will take me several months though.

As for a better amphibious strategy, work on it has started a week ago. Because of point (2), it is now possible to do a serious job on that. Again, it will take time to mature (the first iteration will surely be fire & forget style of amphibious operations, but sustained (à la D-Day) effort can be done eventually... perhaps for VGN)

[tagwyn]

I think the invasion of Penn. in 1863 was done to make the Union Army manuever so Lee could destroy it. Almost worked! Simply beyond an AI, even Athena. T

[GlobalExplorer]

I think in general terms, I'd like Athena to put more weight behind strategic decisions, especially naval invasions. The sea invasion with a lonesome elite brigade is an example how Athena in the end works for the (competent) players.

I already like that it does a lot of unexpected things, but they appear to be of more random nature and not coordinated efforts between larger forces.

I would also like to know if there is a high level strategic layer that (every couple of months) alternates between different modes like

- concentrate on the East / West
- defend in <state>
- build defense line in <objective>
- concentrate behind naval invasion e.g. through Norfolk into Virgina
- rest forces and send raids

etc ..

[Banks6060]

I think we're asking alot of the developers. A "superior AI" experience is FAR off in the future when it comes to games. The best competition, always has been, and I believe always will be a human opponent.

It's asking a lot for guys to program EVERY decision possible into the AI's mind...the ability for the AI to create, develop, perhaps adjust, and carry out long term strategies. It's just not possible IMHO. The AI is reactionary in nature....not pre-meditative.

[MarkShot]

Granted I would say that programming a wargaming AI that passes the "Turing Test" is one of computing great challenges. However, various game engines have made notable progress; already cited is Panther Games. It seems the solution is not in any generic AI approach, but rather programming the type of analytical processes performed by real commanders: analysis of terrain, analysis of road networks, analysis of time, analysis of opposition and force ratios required.

One of the things most impressive about PG's work is that they have a crafted an AI that is capable of producing reasonable grand scale attacks. Since defenses may be static or active, and static defenses are often sufficient when executed under the cloak of player FOW. Defending has usually seen superior results to attacking in games. However, PG's engine can attack. Now, I must also point out that for a game designer to encourage the player to think and play at the strategic level additional measures must be taken besides just writing a good AI. The temptation (especially for John Grog) will always to micro-manage and out perform any AI. So, the game mechanics must: Set up penalties for micro-managing that make it preferential for the player to rely upon proxy control. However, these penalties must be realistic and not arbitrary if the game is to be immersive. PG uses the dimension of time and order delays. A player who attempts to micro-manage everything sees his order delays go through the roof and his force reactions effectively become so slow that any advantage of human planning is negated by sloth. It should be noted that this is not some penalty applied just to make the game system work. In the real world, if one has a big project and then attempts to personally handle every detail instead of delegating, well, then, that commander/manager will become the bottleneck for all action. Everything will grind to a halt. Effective and fast real world command/management structures depend on delegation, limited autonomy, and subordinates interpreting and executing the commander's intent. Bottomline: Order delays do not feel like a gamey restriction, but rather like real life when you break a functioning organizational structure.

I think the current state of dissappoint AI in PC games is more a function of industry focus and behavior. First, I would wager that most mainstream developers set out to produces games that are playable for X months and not necessarily more (unless community content development kicks in). So, if a game is playable 6-12 months already, then it has met its marketting/design requirements. It is more a freak accident, when a game ends up having an AI that offers more than a year worth of challenging play. Such accidents do sometime happen due to individual designer talents, the tenacity of programmers to code what they want despite the pressure of management, designers who are in the business to realize dreams and not profits, ... However, it is rare. More often, limited AI is good enough and, as we know, major studios focus on graphics and the 3D epic spectacle.

Well, I have spoken about three companies/designers here: Panther Games (Dave O'Connor), AGE Studio aka AGEOD (Philippe Thibaut/Philippe Malacher), and Adam Bryant. All of these designers are uniquely passionate about what they do. As Pocus has recently said in one interview yet to be posted (for the Brett Schulte's ACW blog), he isn't simply designing/coding this stuff, but he plays it like the rest of us. The same with Adam and Dave. They love to find time to go off in a corner and play against their engines ... also, Dave's case he loves to watch it play against itself. So, these men are passionate about war gaming and delivering a better experience, since they, themselves, want to see and experience it. Beyond that there are individual factors that affect the way they work.

Dave O'Connor is perfectionist. It is no secret that his games are often way late despite posted dates. He'll focus on real world processes and he will spend any time necessary to get those things right. This is put before any business concerns. He has produced a great engine, but at the same time, wealth has thus alluded him too. There is a price for perfection.

When, I look at AGE Studio, I think one thing that has allowed our AI to get to where it is both the software architecture and company policy. As you will have noticed, many enhancing patches are done upon games that already completed. Effectively, AI improvements are continuously being backported to games which have been out for 6-12 months or more. The architecture facilitates that backporting and company policy is to continue to improve games long after the launch date. So, the AI that BOA will end its software life cycle with is far beyond the AI which it was born with. It's AI has evolved substantially. Effectively, what was budgetted for BOA's AI was acceptable, but ultimately, backporting, allowed BOA to realize an AI way beyond the limited project planning which was part of its software life cycle. Since BOA's evolving AI was due to backporting and work on future games, then the AI going into future games was not simply untried code and with limited beta tested enhancements. Instead the AI going into future games has proven itself in the theater of customer play and expectations before being realized in the latest products. Aside from Thibaut's/Malacher's genius, I think this is another factor that accounts for the quality of AGEOD's overall high standard of AI. That's long term product support and backporting.

[denisonh]

IMO one of the biggest obstacles to improving both the AI and the player management tools is the lack of modeling command and control functions.

What I mean is that the processes the MarkShot has mentioned need to be overlayed in a model that represents the planning staffs. This would have several benefits in that it could be a tool for the player to assist in management of tasks and the AI in supporting processes for making "decisions".

The execution of strategic, campaign and operational orders needs to be modeled to translate the outputs of the staff assisted decision making to translate into action at lower level. "Instantaneous" execution of orders above company level are problematic: getting an Army with multiple corps going with all the coordination and support necessary to be integrated and effective takes quality staff work derived from timely and decisive decision making. The old "1/3 - 2/3" rule is critical in operational planning and representative of the planning tasks that have to performed at every echelon (The rule is if you have 3 days until you have to attack, use 1 day to develop your plan and give your subordinates 2 days to do theirs). This construct allows for planning and adequately supporting the operation from both an operational context and a logistical one. The delay mentioned by MarkShot abstracts this to a certain degree, but does so without really representing the process that "generated the delay". The complexities of campaign/operational planning can be abstracted, but the key inputs, processes, decisions and outputs should be represented.

Most of these functions are "assumed" by the player and are not specifically modeled in computer games, making it a task for the programmer to create an AI to do all these functions. A by product of this "assumption" is a level of complexity that can overwhelm players. Imagine being an Army commander, chief of staff, adjutant, intelligence officer, operations officer and supply officer all at once.

Model the staff and process for transmission and execution of orders. Allow both the player and the AI to focus on decisions.

[Jabberwock]

Athena is a great AI.

That said, a partial reply to part 2 of your question:

Will peform a hasty attack if it perceives that the enemy is disorganized; will perform a careful cohesive attack against a prepared defense.

She is too hasty to exploit perceived opportunities. She lacks both the ability to think turn to turn, and the situational awareness of a human, and that is why it is relatively easy for a human to trap her. I think if she had scripted operational strategies to choose from in appropriate situations, and then had a way to compromise or make a weighted choice between those strategies and the opportunistic "blitz" behavior she often exhibits now, that would improve her performance in the short term.

In the longer term, I think that would delay her development towards passing a Turing test, and eventually achieving world domination.

More specifically, I'd like to see her give more weight to the value of entrenchments both defensively and offensively, and to maintaining cohesive front lines of corps, especially where the terrain is favorable.

[Chaplain Lovejoy]

I notice that there's a check box on the menu to make the AI better by giving her more time. Can the AI be made better yet by giving her even more time than that check box allows?

[MarkShot]

(1) As far as I know, even if you edit the data file, you only have a binary choice. You cannot assign the number of CPU cycles allocated to a computation.

(2) My guess is that the more time to the AI is used to measure the depth of the search space. A chess program will look ahead so many moves and compute all possible outcomes. I would imagine that Pocus is doing something similary. My guess would be that checking that allows the AI to search one more turn into the future or deeper into the tree of possible moves.

So, at the moment, that's the best you can get for increasing difficulty, but for giving the AI better vision (assymetrically lifting the FOW).

Over the long term, the plan is for AGE to be able to plan and fight using longer time frames.

I recently just sent Pocus (and another game designer, Dave O'Connor) an interesting article on emerging technology where the massively parallel computational architecture available on video cards could be applied to enhanced the performance and concurrence of AI processing. NVidia is designing video cards that can handle more than just video and providing a general purpose API for programmers to get at all that extra power. See:

http://www.tomshardware.com/reviews/nvidia-cuda-gpu,1954.html

[Chaplain Lovejoy]

As a lifetime member of the United States Chess Federation, I have followed with interest the attempts to make computers play better against humans. I'm a little rusty on this, but it seems that there have been two general approaches.

One approach was to figure out ways to not allow the computer to waste CPU cycles on obviously ridiculous or suicidal moves. The other approach was more brute force in nature: depend on the ever-increasing prowess of computers to allow as much computational depth as possible on all potential moves.

In games like AACW, there's a random element (dice rolling) that doesn't exist in chess. Perhaps this factor would bog down the brute force approach too much.

[MarkShot]

I had recently read about a team that had developed a cannot lose chess engine. It basically goes like this.

When there are no more than X pieces left on the board the engine has sufficient computational power to computer every move down to the end and therefore play perfect from that point. The approach is basically to play a decent heuristic game until X is reached. At which point, the engine switches to full forward prediction mode and will win. The only chance an opponent would have would be to position the program on a losing path prior to point X and then keep there for the remainder of the game.

As an exercise in inference, then it is clear that program could be written to defeat The Program by being able to play a perfect game from X+1. Which probably leads to a simple proof that a properly programmed machine could win or lose a chess game based on only the very first move.

[pasternakski]

recently just sent Pocus (and another game designer, Dave O'Connor) an interesting article on emerging technology where the massively parallel computational architecture available on video cards could be applied to enhanced the performance and concurrence of AI processing.

This is highly interesting stuff, and thank you for posting the link. The intriguing notion of replacing CPU dynamics with the more advanced techniques becoming available through GPGPU may, someday, save the poor computer wargamer who prefers AI play (and those subjected to the torture of watching horribly flawed AI handling of subordinate functions; see Matrix's UV and WitP as prime examples) from throwing up his hands in frustration and moving on to other pursuits (remember, folks, that three-fourths of computer wargames are played solitaire - a statement backed up by the survey conducted on these forums).

If you lose the battle for that audience, you have lost the war.

So, it is good to see that Nvidia, in its ongoing quest to slay the ATI dragon, has advanced the boulder a few feet farther up the side of the mountain, in a way that may help wargame AI designers to improve their product (and bless those mentioned by Markshot in another post, including Philippe Thibeault, one of my heroes, for their efforts).

The problem, as I see it, is this: who feeds the bulldog? As the author of the linked piece points out, these advances are great for the company making them only if they lead to increased profits and emasculation of the competition in support of realizing the bottom line. What is the advantage for a wargaming company that, as was so astutely pointed out, is interested only in publishing games that provide a few months of playability? If you create a game that wholly absorbs your customer base's attention permanently, how are you going to sell new games in the future (considering that the customer base is a niche one)?

As Klang said in the Beatles movie, "Help!",about trying to sustain himself as head of his Middle Eastern religion, "it's a real problem."

I suggest, at this early stage of the discussion (which has gone on now for some 30 years without palpable results), that the first thing designers have to do is discard the idea of making the AI "think like" a human opponent.

Sure, the resulting AI behavior should create intelligible results. Still (and taking the lesson from chess-playing programs), the technology ought to be invoked on the basis of making the hardware and software work as what they are, not as what naive Internet forum participants think they should be.

Let's leave design to the designers. Me, I prefer gaming. And I still hope for fulfillment of the old, old promise that computer wargames will "...provide a challenge even when no human opponent is available."

Besides, most human opponents turn out being a pain in the @$$.

[chainsaw]

In the longer term, I think that would delay her development towards passing a Turing test, and eventually achieving world domination.

While you're at it give it an Arnold Schwarzenegger accent, leather jacket and a big shotgun and its "Hasta la vista baby"!

Seriously, the 2 points I want improved have been mentioned:
1 - ability to plan and undertake amphib assaults (and then pour in reinforcements and supplies to make it a serious attack); and
2 - defend in depth the most important objectives (DC or Richmond, etc) and not leave it open to go chase a feint attack.

.