Observations on the Tactical AI

[Edwin P.]

Over the past few months I have posted a number of thread regarding the strategic AI. Now its time for a single thread with observations on the tactical AI.

1> Game #1 - Expert - Axis AI

At the start of the war Germany fielded an extrremely large army of armor and infantry units along the Russian-German border (about 20 units in Germany alone); however, the commanders of these units (5 German HQs and and 2 Italian HQ) relaxed in the spas of eastern Romania oblivous to Hitler's forthcoming declaration of war on Russia.

It should be noted that the war with Russia began with 1 German airfleet assigned to the eastern front and 5 German air fleets watching the coast of France.

In spite of the lack of command support the German army managed to inflict huge losses on the defending Russian army, One can only be thankful that the German commanders were not available to coordinate these attacks and that the German pilots were relaxing in France and Northwestern Germany.

Much later, with the Russian army advancing in the east and a D-Day in progress on the west Germany and Italy entrenched 2 HQ, 2 Corps, and 3 (1 Itlian and 2 German) Air Fleets in the mountain region to the east of Switzerland while another 2 German HQ units and 3 air fleets relaxed near Belgium. Too far from the east and west fronts these units made no contribution to the defense of either front. Meanwhile another German HQ relaxed in Northern Italy soaking in the sun.

It should be noted that 1 Italian Corps, 5 German Corps, 3 German Armies and 2 armored units perished defending France without the support of air power or HQ units. The allies can only be thankfull that the German high command did not assign any HQ units or airfleets to support their forces fighting the Allied invasion force and that the Germans responded slowly to the Allied buldup in France.

A quicker response supported by HQs (3 were available in region), Italian infantry units guarding Italy and nearby air fleets (1 Italian and 5 German air) not assigned to the Eastern front would have ment disaster for the American invasion of France.

The allied invasion force was supported by two Str 13 UK fighters in the UK, one 11 Str US Fighter, and one Str 2 French Fighter and two carriers.

--------------------------------------------

Notes: In this game the Axis AI executed the perfect production strategy but failed to commit its HQs or Air Fleets to land battles on the Eastern or the Western Fronts. This was probably caused by the concentration of the British fleet in the Baltic Sea and its success in destroying the German Navy. If the AI had commited these forces to battle the human player would have faced a much tougher opponent.

--------------------------------------------

[ May 12, 2003, 09:50 AM: Message edited by: Edwin P. ]

[Edwin P.]

2> Game 2 - Expert - FOW OFF

The German Armies destroy the corps defending Brussels then 3 German air fleets bomb the defenseless city before the German Army advances in to take it.

The German attack on France continues supported by 3 HQ but no air support for the attack on Paris. Italy subsequently takes Paris after heavy fighting by the German Army defeats the 5 French Corps defending Paris.

A lone Canadian Army unit is left holding Brest, Against fierce never ending German and Italian assaults it survives attack after attack. Reinforcing itself in the breaks between battles it has been beaten down to Str 2 and 3 often but has never fallen. The Canadians are grateful that the German and Italian high commands did not support their attacks with air strikes or a HQ unit.

Following the fall of Paris the Germans DOW on Poland and advance towards the capital with a lone HQ unit, believing the Danes to be cowards at heart. Yet the Danes resist. The Germans bomb the city and a lone UK warship blocks the straits to the capital of Denmark. The bombing of Denmark continues. The HQ unit retreats towards Germany. After three months of fruitless bombing the Germans send increase the number of air fleets assigned to the task and destroy the danish corps before turning their bombers on the UK warship, which is quickly reduced to Str 1 and retreats. The German HQ unit takes the Danish capital 3 turns later.

While 2 Italian ships bombard Malta the Italian Sub crews relax in Bari while the crew of the two Italian battleships have a good time in Venice, although there are no allied ships in the Med, proving the ploint that Italians are lovers not fighters.

Germany DOW on Russia, even before Russia prepares for war, with 3 HQ, 1 Armor, 8 Armies on the Russian border, but no air support. The Axis pilots are relaxing in Rome, Brussels and Paris. The initial attack on Russia is ineffective and the Russians make an orderly withdrawal from the front lines.

When the British invade France with 2 corps to take advantage of the Canadian Army's foothold in Bari the Axis have waiting 3 HQ, 3 Air, and 3 Armies in a line streatching south west from Paris. The only forces remaining in Great Britain are 1 strategic bomber (Str 5), 3 UK airfleets, 1 French Air fleet (Str 4) and a UK HQ unit.

Yet the Germans fear to cross the channel in fear of the UK Navy deployed around the city of Brest and off the shores of Leningrad.

-------------------------------

Notes: In this game the AI used its HQ units properly to support the attack on Russia but failed to properly concentrate its land or air forces to this battle. Instead it kept about 9 infantry units in France to stare down the Brits holding Brest and these units were without HQ support until after the majority of these forces were withdrawn to the Eastern front, after the war with Russia had begun.

Throughout this game the German AI failed to properly use its air fleets in support of its infantry attacks and often deployed them far from the battle lines.

In the case of the Canadian Army holding Brest a human would have attacked Brest with Air and HQ support or stationed a small holding force in France while moving the excess armies to the eastern front to support the attack on Russia. Futhermore, the AI occassionaly abandoned the hex adjacent to Brest, allowing the Canadian unit to be reinforced to a strength of 8.

----------------------------------------------

[ May 12, 2003, 11:40 AM: Message edited by: Edwin P. ]

[Edwin P.]

Game 3> Expert - Axis AI - FOW OFF

The Allies plan to defend France with Armies stationed along the borders. The Germans reply with a DOW on Belgium and the destruction of the Belgium corps north of Brussels. Two turns later the capital city of Belgium falls following an extensive air bombardment and the French fall back to defend Paris abandoning the Maginot Line.

The Germans pursue. The advancing German armies are supported by 3 German HQ units and 4 Air Fleets which advance in the wake of the German advance to give full support to the German Army. The Air Fleets concentrate their fire on one unit at a time and the French are quickly destroyed. The French manage to evacuate 1 Army and 1 corps to England. No human could have done better.

Meanwhile the British Navy has concentrated all of its forces in the Med to combat the Italian Navy, which was quickly destroyed with minimal damage to the British Fleet. Yet you have to admire the bravery of the Italians whose Ships advanced into combat in the face of a air unit on Malta, a bomber unit in Algeria, 3 Carriers, the entire French Fleet, and most of the British Fleet (except for 2 naval units off the port of Manchester).

While the Naval battle was going on East of Malta the Italians sent a corps transport to sieze an undefended Gibraltar. The UK fleet commander seeing this rushed a canadian corps from off the coast of France and a Fleet was diverted from the naval battle to attempt an interception. Alas, the Canadian corps beat the Italian corps to Gibraltor with minutes to spare.

Following the fall of France the AI executed a 2 turn conquest of Denmark supported by German Naval, Air, Armor and HQ units.

Most strangely the Germans in France also made a run for an ungarrisoned Gibraltar. They reached the shores of Gibraltar and were attacked by one UK Battleship. Reduced to a strength of 7 they had a choice, land on Gibraltar and seize the Rock or retreat. The German commander decided to retreat towards America.

The Axis attack on Russia was well planned and supported. With 6HQ, 3 Air, 8 Armor, 16 Armies and 5 corps it was an overwhelming force that could only have been improved with the addition of the the 3 German air fleets stationed in France and the German Armor guarding Belgrade.

---------------------------------------------

Notes: In this game the German AI executed an almost perfect attack on France and Belgium. They could have conquered Belgium in 1 turn instead of 3 if they had focused their initial air and land attacks on Brussels instead of the corps north of the city. As for the Italian fleet, in the face of such overwhelming odds most humans would have moved the Italian navy to the safety of the northern Adriatic Sea where they could be protected until air fleets could be summoned to Italy and the French fleets are disbanded with the Fall of Paris.

Rushing the Italian Corps on Salerno to Gibraltar while the UK fleet was engaged around Bari was quick thinking on the part of the AI. If the Canadian corps was not already off the coast of France the Allied fleet would have been bottled up in the Mediterranean Ocean for the duration of the war.

The AI also took advantage of an ungarrisoned Gibraltar later in the game and reached it shores with a German Army transported from France; however, it decided not to occupy Gibraltar after suffering minor damage from a UK ship that recently arrived in port. A human player would have seized the Rock and thus permanently seperated the Allied Naval fleets.

The German attack on Russia was also much stronger this game; however, the AI still did not advance his HQ units into Russia to support the advancing troops.

-----------------------------------------------

[ May 12, 2003, 08:57 PM: Message edited by: Edwin P. ]

[JerseyJohn]

I don't get it. :confused: Unless you're implying the AI occasionally misses the best possible moves! :eek:

[ May 12, 2003, 01:03 PM: Message edited by: JerseyJohn ]

[Edwin P.]

JerseyJohn

I am trying to highlight ways in which the AI could be improved at the tactical level by sharing my observations on the progress of three games vs the Axis AI during the 1939 Scenario.

The key element thus far seems to be that the AI needs to better concentrate its forces against an objective and support them with HQs and Air Fleets. (ie in attackig Russia the HQ's and air fleets need to continually advance behind the German armies as they push the front forward. All too often the German units push forward while the HQ units remain far behind the front lines.)

Other issues such as bombarding empty city hexes which are about to be occuppied, not taking the Rock, targeting French Ships with Air power as opposed to British ships or French Infantry, and avoiding unnecessary operating of units are also issues for fine tuning of the AI.

[ May 12, 2003, 02:24 PM: Message edited by: Edwin P. ]

[JerseyJohn]

Edwin

Of course, I was just screwing around a little.

The AI comes up with some truly bizarre tactics.

I've seen weaker AIs, mainly twenty years ago on the old Commodore 64 (which also had some pretty good AI games) but I've never seen any with reactions as weird as SC's.

Glad you laid it out in such detail. I think, after the playing map and lack of true weather, the AI has garnered more comment and complaints than any other aspect of the game. I've found, for solitaire purposes, the AI can't handle the AXIS at all; in France I've seen it avoid an open Paris to attack the Maginot Line. Which is not to say it can hanle the Allies either, because it can't.

I don't know what the answer is. No doubt Hubert, he's got to be aware of this glaring weakness, has been working on the problem for SCII.

Regarding AI in general: In the early days of mainframe chess matches, it wasn't uncommon to see two highly touted programs play down to opposing kings, which proceeded to chase one another around the board till the automatic stalemate rule saved the day.

That branch of AI, without dice or probability tables (but also without true intuition on the machine's part) has beed developed to the point where, at present, desktop computers play on an extremely strong master level without any of those early bugs like the one described above.

[ May 12, 2003, 02:55 PM: Message edited by: JerseyJohn ]

[88mm]

I get it. It's a stupid game

[Edwin P.]

JerseyJohn

They key in the chess programs that you mentioned and other games with a good AI is that they are constantly evolving and being updated with new routines and fuzzy logic as the developers and users discover weaknesses in them. This takes lots of time and money, unless it is designed as an open source program (aka Linux) where user groups contribute to the program's development.

That said I think that their are several layers of AI and that some part of the AI can be opened to player involvement. Of course this is just my off the top thoughts on the subject and this would need to be further developed if Hubert decides that this is the path that SC/BattleFront/Fury wants to follow.

Example: Research Strategy, Production Strategy, Opening Moves, Resources required for securing specific objectives, Conquest Roadmap

To elaborate;

1. Research Strategy - A Text File that outlines the research strategy of a nation

UK: 2LongRange,1Jets

US: 2AntiTank

Russia: 2AntiTank,1Jets

2. Production Strategy - An production strategy for utilizing plunder

Example:

German:France:2Air,1Research

German:Poland:1Air

3. Resources Required for Objectives - A text file that gives the AI the desired resources for an objective and the AI executes the plan.

Example: Iraq Hex (xx,yy), 6Land,1HQ,1Air

4. Conquest Roadmap in a text file

Example:

Germany: PolandxDenmark, Belgium, France, Norway, Sweden (where x is same time as)

Italy: IF (ALLIED NAVY in MED = 0) Then Greece > AlgeriaxSouth France > Beruit > Syria > Egypt

[JerseyJohn]

Edwin

The difference also with a game like chess is a huge database of mastergames can be utilized by the AI in deciding things such as openening and ending strategies; in chess the Middlegame involves the greatest amount of original thought. In effect knowledge of opening theory gets you to the best possible middlegame position and knowledge of endgame theory leads you out of the endgame into the best possible conclusion.

Additionally, in chess there is no FOW and no variables of chance etc. . . .. Some in the forum have said chess is a simpler game than SC, which is probably true from the computer's standpoint but most assuredly not from a human player's point of view.

Anyway, the problem is you'd have to assign values to different courses of action and weight values in logical decisions for the computer, which is not beyond the realm of possibility, it's only a matter of how far the programmer wants to take the project. I've never seen a complicated war game where the AI makes very good decisions.

Hopefully Hubert can improve the decision making capabilities of this game, but I think it will be difficult and time consuming in programming terms. Most people have no idea how much information the human mind is capable of running through in a matter of nano-seconds. We call it intuition or various other vague things, but I remember playing against developing programs in the eighties that a human player automatically eliminates hundreds or even thousands of poor possibilities that the computer needs to investigate. It does so very quickly but the sheer volume of moves looked into takes a lot of time.

I suspect something similar can be applied to a war game, but if it is we might be talking about four or five minutes each turn of watching the AI doing nothing because it needs to analyse everything! before arriving at a plan -- or plans, divided into zones on the map tying into a single overall plan.

I don't know whether or not this is how wargame AI is organized, but I do know it's the reason chess programs play a much stronger game on the slower settings. I doubt most war gamers want to sit and stare at the screen while the AI set at Fieldmarshal Difficulty spends ten minutes considering it's moves. Then, because each turn is made up of many different individual moves, it would need additional time to reconsider changes in subsequent piece moves as the situations unfolds with variations to the overall plan it's attempted to put into action.

Which is not to say that in ten years, or even five, there will be machines out there running twenty or a hundred times faster than the best current machines and such an approach will present no problem at all. By then people will choose to play other people for the opposite reason they do so now; it will be so they can have a chance to win a game.

[ May 12, 2003, 08:16 PM: Message edited by: JerseyJohn ]

[Edwin P.]

JerseyJohn

I strongly agree , at the same time I think that SC is similar to Chess in the same manner that you mentioned.

1. There are a limited number of opening game strategies.

- What do you do with the UK Med Fleet?

- Attack Denmark at the same time that you attack Poland?

- How do you attack Belgium if Belgium is still neutral?

- After France, what next?

I think that the SC AI needs a much larger library of opening game moves.

2. The hardest item to program is the middle game.

3. There are a limited number of end game strategies.

- IE What can the US do when France, the UK and the USSR have fallen?

- IE What can the UK do when the USSR has fallen?

- IE What can the Axis do when D-Day strkes in France (Concentrate France / Concentrate Russia / Fight on Two Fronts)

Just once I would like to see the AI respond to an allied D-Day by operating in large number of troops,HQ, Air and crushing the invaders and staging a limited withdrawal in Russia. When the invaders are defeated operating these forces back to the Eastern front.

[ May 12, 2003, 09:03 PM: Message edited by: Edwin P. ]

[Shaka of Carthage]

I am not sure how I want to say what is going thru my mind, so I will just throw it out as it comes to me.

AI is probably the single most difficult thing to deal with in a wargame.

While Mr H, may feel otherwise, almost all of us love SC and offer our "suggestions" only as a means of making it better.

Since all of us want different things, if Mr H attempted to implement all of our suggestions, SC would become an unmanageable mess.

Everything can be done, the only limits being time and money.

AI, what is it to most of us? No matter what you do, the AI can never equal the challenge of playing agianst another human. So should the AI be nothing more than a practice dummy?

Once you have mastered the AI, play TCP. But if you can't do that, how about PBeM? And TCP, PBeM actually works in SC, go figure.

Open Source AI, that would give us the "users" the ability to enhance the AI. Have any of you heard about having too many cooks in the kitchen?

Open source works fine by letting people see the final product and throwing suggestions out for enhancements. Every know and then someone will come up with a simple solution for a problem no one was able to fix. The other way open source works is for each "chef" to be able to create his own little area. But then you need a "master chef" who understands modularity interfacing.

How many of us here understand what "fuzzy" logic is? How many of us here are actually logical?

The person who figures out how to make AI as easy to modify as a scenario editor does, will make a million bucks. But by then, that million will probably only buy a car.

Whoever it was that said the British Commonwealth could have won the war alone, made a good point. Incomplete, but still a good point. Too bad the flag waving overcame logical discussion. But thats another topic, no need to hijack this one.

The SC AI is heads and tails better than what is out there. Then, I have seen and played against some terrible AI. So to me, since I rarely read the manual anymore intially, the AI allows me to start a game, and start over once I realize the dumb stuff I have done. AI is always there for me, even in the middle of the night. And it never complains. And every now and then, it does some things to surprise me. But just like a newbie, it will do some dumb things. However, it is for most of us, just a stepping stone to playing human opponents. Even though some of my Ladder opponents have very little humanity.

You have got to love a guy like Edwin P, who day after day will analyze and quantify what the AI does and then come up with ways to do improve it. Mr H should at least autograph a copy of SC and mail it to you. If he really wanted to take care of you, he could send you a gold membership card for Deja Vu.

Some new guy, who hasn't posted much, hit it dead on. The more you read about things, the more you realize you know very little about it. All those "facts" we knew 10 years ago, are washed away by the new information that is found out as time goes by.

Isn't it strange how those who were there, know so little about what really went on? Kinda goes to show you that politicians are a breed all by themselves.

[Edwin P.]

I agree, all excellent points.

As you said, if Hubert implemented all of the ideas he's heard he would surely have a mess on his hands. I trust that he will draw from this forum selected ideas to make for a better SC2.