Turing's Battle

by Michael Wild

The simulation's graphics are elementary.

From what Alfred could tell, this was a more table-driven and list game more akin to an Excel spreadsheet than World of Warships.  But, he was told, it made surprisingly good moves and could be easily programmed with various scenarios from WW2 to the late Cold War and maybe some things from more current.  "It," Alfred ruminated, was a new artificial intelligence naval opponent.

The basic structure, he was told, allows working with a cutting-edge AI.  The table-based game and its turn structure were made to enable the computer folks to build a perfect opponent.  Amazing graphics like World of Warships could follow once the AI could win game after game.  Alfred is a historian and game player - tabletop, not computer - and had been asked to play against the AI.

Alfred smiled at the names of his ships for this scenario, "Not Midway," all South American heroes and locations.  He knows, being a historian, that in the past, the Navy exercise fleets were called Red and Blue, using the names of the existing ships.  "But this is a fictional construct, and instead of using any real names and upsetting a government or U.S. naval folks when the inevitable leaks happen, they picked a South American mix," explained one of the designers of the game.  A young kid in a dark t-shirt and shorts.  Thus, Alfred found he had three aircraft carriers in his fictional fleet: Rio de Janeiro, Benito Juárez, and Simón Bolívar for CV-1, CV-2, and CV-3.  All conventionally powered and carrying less aircraft and firepower than a U.S.-styled Nimitz- or Gerald R. Ford-class carrier.  They seemed to correspond to certain China-built carriers, Alfred observed.

Alfred, running queries, can see the basic breakdown of these ships now reduced to just a few factors.  He has many cruiser-sized destroyers, regular destroyers, some missile and other anti-submarine focused.  In addition, there is a collection of submarines, some useless pre-Cold War copies of U-boats, some conventional Cold War models, and some surprisingly powerful nuclear submarines, Argentina class - likely a copy of Soviet Cold War boats.

The scenario allows him, the aggressor, to pick his target for victory.  He is to pick Alaska or Hawai'i.  Alfred thought, "'Not Midway,' indeed."  His terminal shows a list of his forces: most of his ships were conventionally powered; for the short time that the scenario covers, he was told he did not have to consider the placement of oilers or resupply his warships.

"Let's go alternative history," Alfred says to himself.  He creates a task force with fast cruisers and missile destroyers with ship-to-ground weapons, with the conventional carrier Rio de Janeiro as the focus.  He heads them to a point near Hawai'i, where the 1941 attack was launched.  "If you are laying bait, it is best to be obvious," Alfred says gleefully to his terminal.  "And like any WW2 aggressive plan from 'South America,' I will split my forces," he says to himself, creating another task force and pointing it at the alternative-history selected victory objective, Alaska.  He establishes two task forces of submarines as a vanguard for his Hawai'i attack.  The lesser submarines will arrive before his forces and will likely get noticed.  The small quantity of nuclear subs is sent out early and are in silent (and slow) mode when approaching Hawai'i to intercept any defenders.  He placed them north of Hawai'i near the original American position for Midway in 1942.  "Should always cover the obvious," he says to himself.

"Now for the real attack," he says.  Benito Juárez repeats the Midway plan and attacks small targets like Dutch Harbor with planes and some missiles.  Seward is a tourist town, but it will get some morale-busting attention.  Next, Simón Bolívar will take on the air force bases, destroy the Alaska oil pipeline, and prepare for the land invasion of the Anchorage area.  "Midway has moved north and is now the Battle for Alaska," Alfred says.  "Try that out, AI," he thinks to himself as he sends out his last commands to start his part of the game.

Tim is a soon-to-retire commander in the U.S. Navy.  He has been voluntold to support and participate in the new AI program.  He has been promoted to virtual flag rank in the scenarios to command fictional U.S. Navy fleets.  He is currently playing the "Not Midway" game against an aggressive "South American" AI opponent.  Tim is overseeing a substantial virtual force - way above his pay grade, but he thinks, "What the heck, I am retiring soon - let's do this."

He reviews his forces online, and the charts and processes remind him of the old Avalon Hill board game Midway he used to play against his dad.  The warships and classes of ships have been renamed, and the capabilities are summarized to just a few essential values.  His first responsibility for the virtual command is to arrange his forces in response to some basic intelligence.  The enemy will likely attack Hawai'i and Alaska.  He notes that the enemy forces include three conventional carriers that match a particular China carrier.  The briefing includes his goal: If the enemy's target should become unprotected, the enemy will win, and the land will be occupied.

"We will not be repeating history today, I see," he says to himself.  He arranges his four nuclear aircraft carriers into two task forces with two carriers together in each.  The names of his ships are American but not current.  Lexington is a class of carriers resembling a Cold War version of the Nimitz.  He has Lexington, Saratoga, Ranger, and United States as his main force, CVA-1 through CVA-4.  He has ten destroyers, standard multi-use expensive machines, typical of the current American design for destroyers, here named the Shark class with just numbers, DD-1 through DD-10.

"Concentration is the best defense against an aggressive enemy - Lee did not enjoy attacking a reinforced position at Gettysburg," Tim thinks.  So he splits all the destroyers between the two task forces.  Tim also has some Wolverine class nuclear attack submarines - another name not in use - which he is tempted to send on a hunt-and-destroy mission.  Instead attaches them defensively to the same task forces - these resembling the Los Angeles-class Cold War versions.  "If you concentrate your forces, then do that," he reminds himself.

Where to deploy is the next question.  Tim remembers that the Americans deployed north to have some coverage of Alaska in the original Midway battle.  Tim smiles and sends his ship south and around Hawai'i.  Remembering his father's tricks in the old board game.  "Never go where the history books tell you - someone has read that book too," was his father's explanation after a bad moment for the then young Tim playing the Japanese forces in the out-of-print game.  "One more time, dad, for old-time's sake," he says out loud as if talking to his father.

Tim, now with a plan, reviews his setup and plans.  He set his task force to run quietly and to zigzag.  This is not to avoid torpedo attacks like in the World Wars.  Instead, he orders the zigzag to prevent his task force's wake from appearing on a satellite, giving away his position, speed, and direction.  The order is done by just a setting on his task forces he observes and costs 20 percent of the speed.  "Worth it, I think," he says as if talking to his dad.

Alan is the computer scientist or the newly created title data scientist, and he is watching the scenario named suggestively "Not Midway."  He has two human players thinking they are playing not a human but a cold-hearted and possibly incompetent artificial intelligence simulation.  Alan is calling it a Reverse Turing.  He remembers that another Alan, Alan Turing, famously imagined a day when a computer's response would be indistinguishable from a human's response; this is called the Turing test.  Today, his human players thought the computer was an AI and not a human.  Thus, they would make, intentionally or not, assumptions about their opponent.  As a result, they would not try to outguess or out bluff a human but instead, try to outplay the computer opponent.

Alan is tracing the players' actions.  He has many values he is generating from the game, such as force structure, amount of information available to the players, aggressiveness, and how much information is provided to the players.  These will all become parameters in a set of equations that Alan might use a simple regression process to solve.  Or he and his colleagues will use machine learning processes to determine the best parameters to approximate intelligence.

"Yes, we are reverse engineering a historian and a U.S. naval officer," he says to his team members, who mostly ignore him and watch their numbers and refine predictive algorithms to reproduce the same results they see from the players.  There are hushed conversations as ideas are discussed, quickly coded, and intensely watched.  The mostly introverted computer and data scientists are having a blast; there is barely any noise except keyboard clicks.

Alan and his staff are watching as the action begins to heat up.

The American forces discover the limits of the "South American Alliance" submarine picket line.  Alfred is shocked by the losses as the enemy's high-quality destroyers and attack submarines simply remove Alfred's old-styled forces.  Alfred is also surprised that a game without graphics could produce such an emotional response.  The ship losses are harshly listed on the screen, but the death counts, even virtual, are disturbing.  Alfred feels like he has failed the folks and repurposes the nuclear attack submarines, currently in the wrong place, to head south on a hunt-and-destroy mission.

Tim was at first suspicious when he received the first messages of the detection of conventional submarines by his listening destroyers and attack submarines.  Still, the signals got stronger, and he decided to act, entering commands into his virtual terminal.  In the resulting action, he took no losses but, like Alfred, was surprised by the estimated death counts.  Not something he was used to seeing; even for an enemy, it was disturbing.

Tim changed his settings for his task forces.  "Time to rush and search," he said to his screen.

Next, Alfred launches his attacks on Hawai'i.  Within a few moments, he was getting damage reports from his attack.  The air strike from his carrier and the missiles destroys most of the airpower in Hawai'i.  The attack cost was a third of his attacking force, but the missiles and the going in 100 percent on the air strike worked.  "If you are going, go big," he said.  "Also, I am not going for a second round.  Time to run away," he whispered to his computer.  He orders the task force to cycle the planes to defensive use and to leave.

And that might have worked, but Tim launched his attack the moment he could and let the pilots and technology direct the battle.  "Planes on the deck are targets," he says to himself.  With all the noise of the enemy's attack on Hawai'i, Tim was soon told that his forces had located the "South American" forces.  The enemy aircraft carrier had just landed the planes and only rearmed a few planes and got them in the air before the full attack hits.  The missile destroyers did stop some of the attacks, but the main target, the carrier, took most of the attack.  Some of the destroyers were also taken out.  Tim took 15 percent losses.

Alfred was not surprised that the Hawai'ian attack was a success and that the main force was destroyed; a message on his terminal said that Rio de Janeiro was a sinking wreck.  But his other carrier strike force started on Alaska, and he knew that the virtual U.S. forces must rush to Alaska.  Alfred learned from messages that all four enemy (Tim's) carriers were in a pair of close-moving groups.  Alfred had his quiet submarines headed into action.  "That will be interesting," he was thinking.  "It's a trap," he says.

Alan is watching the action and collecting information on how the players react.  He is building a model of information on how an AI should respond when supplied with certain facts.  His human players provided interesting factors for his models.  For example, he noticed that the players' aggression changed with information and threat size.

Alan was trying to discover parameters and turn them into a single value.  Like chess programs, he thought to himself, you can measure the King's position and give better values for being moved away from the center - it is a simple but effective measure.  The opposite for major pieces - queen, bishop, and rook - which are more effective towards the center.  For a basic chess program, you can just use these calculated values to evaluate legal moves and play basic chess.  Likewise, Alan could already see that if he could quantify a value for information and threat size, he could create an equation of aggression.

Alan is already writing a bit of Python code to crunch the data into some linear regression, and even a machine learning model using the random forest of trees approach to see if he can predict the moves that have already happened.  Other computer scientists and data scientists are heads down and happy.  There is good data coming in!

Tim is not surprised when the United States is torpedoed at full speed and then hit with missiles when stopped.  A quiet nuclear attack submarine, obviously of Soviet design, has slipped by and killed the carrier.  A U.S. carrier is hard to sink, even a virtual one - Tim thinks and smiles grimly.  The United States is now running at 20 percent.  The AI bushwhacked him.

Alan and his fellow scientists watch for how Tim and Alfred react.

Alfred yells, "Yes!" and fist pumps when he sees that the U.S.  AI has lost a carrier.  Not a flaming wreck, but put out of action.  Alfred makes no changes, "Let her ride," he says to himself.

Tim splits his groups apart as they rush to Alaska to stop the incoming invasion.  He leaves two destroyers to protect and help the United States back to port as he remembers that the game counts causalities.  Tim recalls that in World War One the British lost two ships when the U-boat torpedoed the rescue ships.  He is hoping the AI is not going for the WW1 model.

Tim's remaining destroyers are in full ASW mode, and many enemy submarines are now falling to their attacks.  No more attacks reach his three remaining carriers, but the fight has slowed his advance to Alaska.

The attack on Alaska is not enough for Alfred to win - he dearly misses the third carrier now.  He must risk a more decisive attack and moves both virtual carriers closer to the Anchorage area.  Another attack will destroy the remaining opposition and temporarily destroy the oil capabilities of the area.  He needs to finish this before the AI stops fighting down south and decides to head north.

Tim shouts, "Damn the torpedoes," and virtually throws his forces into a mad rush north.  As soon as he reaches maximum range minus his continued steaming speed, he launches another blind attack into the north.  Moments later, the Saratoga, like her WW2 namesake, seems to attract torpedo attacks, but U.S. carriers can have a lot of holes in them, and she keeps going with two less important hits.  Even in a rush, Tim slows the task forces by 10 percent to match Saratoga's reduced speed.  This change also allows for more ASW work, and his display shows multiple losses to the enemy soon.

Tim blindly launched an attack, relying on the same tactic as before; this time, it was less effective, and only some of his forces found the enemy.  The attack hits one carrier, Benito Juárez, which is heavily damaged.

Alfred cannot win.  While he withdraws, Alfred's virtual forces fight off some missile attacks.  Alfred could get the planes off the Benito Juárez and onto the Simón Bolívar using his undamaged carrier and thus can assign all his aircraft to a defensive role.  He is making a fighting retreat but knows the invasion is off.  Alfred will try to exit with as much of his forces as possible.  "Next time, we will get ya," he says in his best Captain Ahab accent.

A lucky hit on Lexington and again on Saratoga reduces the final punch, and Alfred's forces can limp away with both conventional carriers.  Tim saves Anchorage from the invasion.  The simulation ends.

Alan and all the other scientists are quiet and busy crunching the numbers.  They will soon have equations and data models to test against the collected data.  They can then remodel the same play and replay the actions of each human player.

Epilogue

Alan was happy that Tim and Alfred would be joining them again.  This time it will be "Not Coral Sea," and each player will have an AI advisor and virtual fleet.  The initial equations and machine learning models are ready to advise each player.  The U.S. player will have an advisor, Layton, and the South American player will have one named Nelson.  Again, Alfred and Tim would think they are playing an AI, which would be partially true this time.  Alan and his crew were looking forward to more data.

Alan knows it will take a while, but the final goal is a predictive engine that can provide input into the decisions of a U.S. Naval commander so good that you can't tell it is not human.  Turing goes to war in a new age.