ArtificiaI Intelligence in SC2′s ‘Assault on Communism’ and ‘Assault on Democracy’
This article by Alvaro (Al) Sousa will examine the artificial intelligence, or AI, improvements made for the two scenario packs Assault on Democracy (AoD) and Assault on Communism (AoC) for Strategic Command 2—Brute Force.
Fury Software took me on as a new designer for both these packs due to my AI work with the popular mod Brute Force 1939. This was the first global World War II game in which the computer opponent had 16 different grand strategies to win the game for both the Axis and Allied sides. There I showed how it was possible to make an AI unpredictable and reactive at the same time. With the lessons learned from that mod and working together with Fury Software we have developed a stronger AI to challenge players for games to come, including Strategic Command 3.
The ultimate goal of creating artificial intelligence is the development of computer systems to perform tasks and behaviors the same way humans do, but at a much faster rate. Creating true artificial intelligence has proven very difficult. For one thing, human beings are limited in how fast we think, so building an intelligence equal to ours is a formidable challenge. Another obstacle is how the human brain works compared to a computer. Our memory is three-dimensional and dynamic: one cell can connect to many different cells to create memory and thought. Computer memory is linear in that a single memory cell connects only to the one behind it and the one in front of it. With all these barriers to overcome how can a game designer incorporate a challenging computer opponent?
Computers can beat humans in games of absolute information such as chess and checkers. In any game that relies on pure mathematics to win, the AI problem can be solved by using math to make the computer unbeatable. It is far more difficult to develop an AI for a game of incomplete information with complex and changing dynamics—for example, a global conflict game with various and different opponents or even a no-limit, hold’em poker game which has simple rules. But each year game developers strive to learn from previous models to improve the game AI’s performance.
At the core of Strategic Command’s intrinsic AI is a system of algorithms that allow the computer to make tactical decisions. These decisions are based on its current situation within its area calculating for position, supply, and combat values. All decisions are based on a plan set forth by the AI or scripts for a group of units. The game designer uses this tactical fuzzy logic to tell the game engine what strategies to pursue. It is an elegant system that takes the worry of micromanaging battles away from the designer and lets him effectively design the four-star general for the AI.
Over the years this system has been improved through the design of better algorithms and scripting tools. These advances allow the designer to have the AI do whatever he needs it to do. Combined with new scripting techniques and features available to the designer, the two new expansions present a better challenge to players on the tactical and strategic levels.
Assault on Communism has the largest open field map ever designed for a Strategic Command (SC) series game. It is the Eastern Front on a 15-mile-per-square scale. We had to change the AI to handle more units and a larger dynamic front. Programming was improved to help the AI realize better combat opportunities and defensive situations. Improvements were made to protect its important assets such as headquarters and planes. It now chooses stronger units to perform attacks, including swapping units to destroy a target rather than just damage it. It groups support assets for offensives more efficiently to give it that extra strength needed to accomplish a goal. Defensively it uses terrain and analyzes enemy movement to prevent destruction and encirclement. It analyzes when the situation has changed so it can resume going on the offensive. So on a tactical level the AI is meaner and tougher than ever before.
On the strategic level there are two systems in place to assist the AI in challenging players. The first is an operational objective list in which the designer tells the AI what it can attack on its own. This serves as a default offensive path for the AI to follow. We shape the strategic objective list to secure important locations for logistics and positioning. This is the backbone offensive below the scripts. The second system utilizes scripts the designer creates to manipulate strategic goals based on game situations. This is the jewel of the Strategic Command AI that separates this game from others in how it performs. The combination of the intrinsic tactical logic, the system of operational objectives, and strategic-goal control make Strategic Command a powerful game engine.
Many older games focus on making the AI playable. In AoC / AoD we worked on making the AI formidable. This was done by play testing with some of the best wargamers we have known implementing real strategic concepts for war and manipulation of the SC engine. For each scenario we built a winning overall strategy for the AI to defeat its opponents based on war theory and historical hindsight. While in some smaller scenarios the strategic options are limited, favoring the intrinsic tactical AI, in larger scenarios strategies were improved from many dozens of test games. Below are two examples of improved AI we inserted in AoC and AoD.
With “Barbarossa 1941″ in AoC I developed a special defensive script for the AI to fall back based on terrain, positioning, delay, and withdrawal. I used a model from ancient times called the “infantry square” to accomplish this. The AI attempts to keep its most valuable counterattack resources, the tank group, in the rear while infantry-class units hold the front lines behind proper defenses. This model has been quite effective in preventing AI units from getting captured and destroyed. It also allows them to counterattack when their numbers are sufficient.
For offensive power the AI waits until it has enough forces mathematically to go after the enemy. We went one step beyond that for our Eastern Front campaigns. If the AI is limited in its resources it masses the best of those resources and chooses where to attack instead of just attacking everywhere. As the Germans, it makes a choice based on what objectives it holds. Will it attack Moscow, Leningrad, or Stalingrad in 1942 after its initial invasion? On which objective will it focus its firepower? The human player doesn’t know, just as the Soviets didn’t know at the time. The Soviet AI employs counterattacks at strategic positions, and those are also determined based on the situation in the battlefield so the German player can’t predict where they will originate.
As the game gets into the later years the German AI must make a decision on whether to go for the win or realize it must hold what it has and prevent a complete loss, which amounts to a minor German victory. It chooses whether to push its advantage or retreat in 1943, playing a mobile German defense. Our goal is to provide players with a challenge to overcome while having them learn how the strategies work.
In AoD one of the scenarios is called “Flatiron 1943.” This is a hypothetical scenario in which a July 1940 Operation Sealion (invasion of Britain) was a success. The Allies shifted many resources from the Pacific to the European theatre in an effort to reclaim the British Isles and prepare to invade France in 1944 before the Soviets fall. In this scenario we had different challenges to overcome. “Flatiron” involves a strong naval component, as the Allies must occupy Ireland, liberate England, and then finally invade France in 1944. The first goal of the Allies is to take back Ireland and England in 1943. The second goal is to invade France in 1944 and advance to the Rhine River. One problem that develops from a static, predictable AI is that eventually a human player will learn its strategies. In “Flatiron” we altered this one-dimensional strategy to a calculated unpredictable invasion. This was designed by allowing the AI to invade anywhere along the coast . The invasion points were challenges that needed to be solved so that the human player wouldn’t have only one solution to play against the computer.
When the Allied AI plans an invasion it actually has several choices of landing zones that best suit its objectives. Since the Germans can’t be everywhere at once, scripts were written to figure out the best locations to invade based on terrain, ports, airbases, logistics, and a little randomness. That way the AI has the best chance of a successful invasion. It will even land in multiple places at once if the Germans don’t properly defend the coast. When the final invasion of France begins, a human player will not know where the AI is coming from.
For the Germans the challenge was different. By the scenario’s design the defense must be balanced between England and France to make sure the Allies don’t bog down the Axis in England and do a runaround invasion in France. What’s the secret to defending both? Well, you will have to play the game and find out. Instead, I will focus on something else more familiar that everyone understands: D-Day in France.
In 1944 there were conflicting strategies between the German generals on how to properly defend France. Rommel wanted to place armored units near various landing zones to attack the Allied invasion immediately and kick them off the coast. General Leo Geyr von Schweppenburg, who controlled Panzer Group West, wanted to place all the tank formations in a central location such as Paris to attack the enemy in mass. Hitler’s decision was a compromise between both strategies, making neither effective.
For the Germans we decided either plan might be the most effective strategy for defeating the Allies on the beaches; thus, we randomly execute one of them. This keeps the AI’s strategy random, simulating a human opponent. If the Allies do succeed in getting ashore in large numbers we use the same successful defensive scripting method as we did for Barbarossa. Overall, we combine a multiple strategies with randomness, letting the intrinsic AI handle the tactical situations, which should challenge players.
Writing AI for a computer game is very difficult. An extraordinary amount of factors are involved for a game system such as Strategic Command. The designer must work with the programmer to put into effect many algorithms not only to facilitate the smaller functions of the AI but also to incorporate larger ones. Each game has to balance strategic, operational, and tactical goals within the engine. Each scenario must be evaluated with seasoned beta testers who have knowledge of how to implement proven strategy doctrine for that specific campaign. In other words, they literally have to beat the crap out of the AI and exploit every vulnerability until the designer corrects any flaws.
Developing a solid AI takes time and the SC engine is an example of a game in which each incarnation presents a more powerful AI game system to continually challenge the players. One thing to keep in mind is that no AI playing a complex grand strategy game of World War II can play as well as a competent human opponent—well, at least not yet. Artificial Intelligence lacks the ability to learn and to adapt as well as the human brain does but can still provide tough competition. I have played computer war games since the 1980s and have seen the AI evolution from its infancy. I was just a gamer when I thought Strategic Command had an excellent AI system. After working for them and being part of the game’s development, I believe it is the most comprehensive, intelligent, and flexible AI system yet developed.
About the Author
Alvaro (Al) Sousa has been a lifetime gamer since being introduced to chess at the age of 5. At 15, he played his first complex strategy game, Third Reich, and has been hooked on World War II history and games ever since, favoring grand strategy games that generate multiple aspects of play. He is also a successful poker player—you have been warned. Currently, he works for Fury Software as a scenario designer. He says his wife has always told him “do what you love,” and he hopes to follow his passion and make the leap to full time game design work in the near future. His previous work for Armchair General was an article on strategy tips for Strategic Command 2 Gold.