What is the best Battleship AI?
Battleship!
Back in 2003 (when I was 17), I competed in a Battleship AI coding competition. Even though I lost that tournament, I had a lot of fun and learned a lot from it.
Now, I would like to resurrect this competition, in the search of the best battleship AI.
Here is the framework, now hosted on Bitbucket.
The winner will be awarded +450 reputation! The competition will be held starting on the 17th of November, 2009. No entries or edits later than zero-hour on the 17th will be accepted. (Central Standard Time) Submit your entries early, so you don't miss your opportunity!
To keep this OBJECTIVE, please follow the spirit of the competition.
Rules of the game:
- The game is be played on a 10x10 grid.
- Each competitor will place each of 5 ships (of lengths 2, 3, 3, 4, 5) on their grid.
- No ships may overlap, but they may be adjacent.
- The competitors then take turns firing single shots at their opponent.
- A variation on the game allows firing multiple shots per volley, one for each surviving ship.
- The opponent will notify the competitor if the shot sinks, hits, or misses.
- Game play ends when all of the ships of any one player are sunk.
Rules of the competition:
- The spirit of the competition is to find the best Battleship algorithm.
- Anything that is deemed against the spirit of the competition will be grounds for disqualification.
- Interfering with an opponent is against the spirit of the competition.
- Multithreading may be used under the following restrictions:
- No more than one thread may be running while it is not your turn. (Though, any number of threads may be in a "Suspended" state).
- No thread may run at a priority other than "Normal".
- Given the above two restrictions, you will be guaranteed at least 3 dedicated CPU cores during your turn.
- A limit of 1 second of CPU time per game is allotted to each competitor on the primary thread.
- Running out of time results in losing the current game.
- Any unhandled exception will result in losing the current game.
- Network access and disk access is allowed, but you may find the time restrictions fairly prohibitive. However, a few set-up and tear-down methods have been added to alleviate the time strain.
- Code should be posted on stack overflow as an answer, or, if too large, linked.
- Max total size (un-compressed) of an entry is 1 MB.
- Officially, .Net 2.0 / 3.5 is the only framework requirement.
- Your entry must implement the IBattleshipOpponent interface.
Scoring:
- Best 51 games out of 101 games is the winner of a match.
- All competitors will play matched against each other, round-robin style.
- The best half of the competitors will then play a double-elimination tournament to determine the winner. (Smallest power of two that is greater than or equal to half, actually.)
- I will be using the TournamentApi framework for the tournament.
- The results will be posted here.
- If you submit more than one entry, only your best-scoring entry is eligible for the double-elim.
Good luck! Have fun!
EDIT 1:
Thanks to Freed, who has found an error in the Ship.IsValid function. It has been fixed. Please download the updated version of the framework.
EDIT 2:
Since there has been significant interest in persisting stats to disk and such, I have added a few non-timed set-up and tear-down events that should provide the required functionality. This is a semi-breaking change. That is to say: the interface has been modified to add functions, but no body is required for them. Please download the updated version of the framework.
EDIT 3:
Bug Fix 1: GameWon and GameLost were only getting called in the case of a time out.
Bug Fix 2: If an engine was timing out every game, the competition would never end.
Please download the updated version of the framework.
EDIT 4:
Tournament Results:
I second the motion to do a lot more games per match. Doing 50 games is just flipping a coin. I needed to do 1000 games to get any reasonable distinction between test algorithms.
Download Dreadnought 1.2.
Strategies:
keep track of all possible positions for ships that have >0 hits. The list never gets bigger than ~30K so it can be kept exactly, unlike the list of all possible positions for all ships (which is very large).
The GetShot algorithm has two parts, one which generates random shots and the other which tries to finish sinking an already hit ship. We do random shots if there is a possible position (from the list above) in which all hit ships are sunk. Otherwise, we try to finish sinking a ship by picking a location to shoot at which eliminates the most possible positions (weighted).
For random shots, compute best location to shoot based on the likelihood of one of the unsunk ships overlapping the location.
adaptive algorithm which places ships in locations where the opponent is statistically less likely to shoot.
adaptive algorithm which prefers to shoot at locations where the opponent is statistically more likely to place his ships.
place ships mostly not touching each other.
Here is my entry! (The most naive solution possible)
"Random 1.1"
namespace Battleship
{
using System;
using System.Collections.ObjectModel;
using System.Drawing;
public class RandomOpponent : IBattleshipOpponent
{
public string Name { get { return "Random"; } }
public Version Version { get { return this.version; } }
Random rand = new Random();
Version version = new Version(1, 1);
Size gameSize;
public void NewGame(Size size, TimeSpan timeSpan)
{
this.gameSize = size;
}
public void PlaceShips(ReadOnlyCollection<Ship> ships)
{
foreach (Ship s in ships)
{
s.Place(
new Point(
rand.Next(this.gameSize.Width),
rand.Next(this.gameSize.Height)),
(ShipOrientation)rand.Next(2));
}
}
public Point GetShot()
{
return new Point(
rand.Next(this.gameSize.Width),
rand.Next(this.gameSize.Height));
}
public void NewMatch(string opponent) { }
public void OpponentShot(Point shot) { }
public void ShotHit(Point shot, bool sunk) { }
public void ShotMiss(Point shot) { }
public void GameWon() { }
public void GameLost() { }
public void MatchOver() { }
}
}
Here's an opponent for people to play against:
- http://natekohl.net/files/FarnsworthOpponent.cs
Instead of using a fixed geometry-inspired strategy, I thought it would be interesting to attempt to estimate the underlying probabilities that any particular unexplored space holds a ship.
To do this right, you'd explore all possible configurations of ships that fit your current view of the world, and then compute probabilities based on those configurations. You could think of it like exploring a tree:
an expansion of possible battleship states http://natekohl.net/media/battleship-tree.png
After considering all leaves of that tree that jive with what you know about the world (e.g. ships can't overlap, all hit squares must be ships, etc.) you can count how often ships occur at each unexplored position to estimate the likelihood that a ship is sitting there.
This can be visualized as a heat map, where hot spots are more likely to contain ships:
a heat map of probabilities for each unexplored position http://natekohl.net/media/battleship-probs.png
One thing I like about this Battleship competition is that the tree above is almost small enough to brute-force this kind of algorithm. If there are ~150 possible positions for each of the 5 ships, that's 1505 = 75 billion possibilities. And that number only gets smaller, especially if you can eliminate whole ships.
The opponent that I linked to above doesn't explore the whole tree; 75 billion is still to big to get in under a second. It does attempt to estimate these probabilities, though, with the help of a few heuristics.
Not a fully fledged answer but there seems little point cluttering the real answers with code that is common. I thus present some extensions/general classes in the spirit of open source. If you use these then please change the namespace or trying to compile everything into one dll isn't going to work.
BoardView lets you easily work with an annotated board.
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Drawing;
using System.IO;
namespace Battleship.ShuggyCoUk
{
public enum Compass
{
North,East,South,West
}
class Cell<T>
{
private readonly BoardView<T> view;
public readonly int X;
public readonly int Y;
public T Data;
public double Bias { get; set; }
public Cell(BoardView<T> view, int x, int y)
{
this.view = view; this.X = x; this.Y = y; this.Bias = 1.0;
}
public Point Location
{
get { return new Point(X, Y); }
}
public IEnumerable<U> FoldAll<U>(U acc, Func<Cell<T>, U, U> trip)
{
return new[] { Compass.North, Compass.East, Compass.South, Compass.West }
.Select(x => FoldLine(x, acc, trip));
}
public U FoldLine<U>(Compass direction, U acc, Func<Cell<T>, U, U> trip)
{
var cell = this;
while (true)
{
switch (direction)
{
case Compass.North:
cell = cell.North; break;
case Compass.East:
cell = cell.East; break;
case Compass.South:
cell = cell.South; break;
case Compass.West:
cell = cell.West; break;
}
if (cell == null)
return acc;
acc = trip(cell, acc);
}
}
public Cell<T> North
{
get { return view.SafeLookup(X, Y - 1); }
}
public Cell<T> South
{
get { return view.SafeLookup(X, Y + 1); }
}
public Cell<T> East
{
get { return view.SafeLookup(X+1, Y); }
}
public Cell<T> West
{
get { return view.SafeLookup(X-1, Y); }
}
public IEnumerable<Cell<T>> Neighbours()
{
if (North != null)
yield return North;
if (South != null)
yield return South;
if (East != null)
yield return East;
if (West != null)
yield return West;
}
}
class BoardView<T> : IEnumerable<Cell<T>>
{
public readonly Size Size;
private readonly int Columns;
private readonly int Rows;
private Cell<T>[] history;
public BoardView(Size size)
{
this.Size = size;
Columns = size.Width;
Rows = size.Height;
this.history = new Cell<T>[Columns * Rows];
for (int y = 0; y < Rows; y++)
{
for (int x = 0; x < Rows; x++)
history[x + y * Columns] = new Cell<T>(this, x, y);
}
}
public T this[int x, int y]
{
get { return history[x + y * Columns].Data; }
set { history[x + y * Columns].Data = value; }
}
public T this[Point p]
{
get { return history[SafeCalc(p.X, p.Y, true)].Data; }
set { this.history[SafeCalc(p.X, p.Y, true)].Data = value; }
}
private int SafeCalc(int x, int y, bool throwIfIllegal)
{
if (x < 0 || y < 0 || x >= Columns || y >= Rows)
{ if (throwIfIllegal)
throw new ArgumentOutOfRangeException("["+x+","+y+"]");
else
return -1;
}
return x + y * Columns;
}
public void Set(T data)
{
foreach (var cell in this.history)
cell.Data = data;
}
public Cell<T> SafeLookup(int x, int y)
{
int index = SafeCalc(x, y, false);
if (index < 0)
return null;
return history[index];
}
#region IEnumerable<Cell<T>> Members
public IEnumerator<Cell<T>> GetEnumerator()
{
foreach (var cell in this.history)
yield return cell;
}
System.Collections.IEnumerator System.Collections.IEnumerable.GetEnumerator()
{
return this.GetEnumerator();
}
public BoardView<U> Transform<U>(Func<T, U> transform)
{
var result = new BoardView<U>(new Size(Columns, Rows));
for (int y = 0; y < Rows; y++)
{
for (int x = 0; x < Columns; x++)
{
result[x,y] = transform(this[x, y]);
}
}
return result;
}
public void WriteAsGrid(TextWriter w)
{
WriteAsGrid(w, "{0}");
}
public void WriteAsGrid(TextWriter w, string format)
{
WriteAsGrid(w, x => string.Format(format, x.Data));
}
public void WriteAsGrid(TextWriter w, Func<Cell<T>,string> perCell)
{
for (int y = 0; y < Rows; y++)
{
for (int x = 0; x < Columns; x++)
{
if (x != 0)
w.Write(",");
w.Write(perCell(this.SafeLookup(x, y)));
}
w.WriteLine();
}
}
#endregion
}
}
Some extensions, some of this duplicates functionality in the main framework but should really be done by you.
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Drawing;
using System.Collections.ObjectModel;
namespace Battleship.ShuggyCoUk
{
public static class Extensions
{
public static bool IsIn(this Point p, Size size)
{
return p.X >= 0 && p.Y >= 0 && p.X < size.Width && p.Y < size.Height;
}
public static bool IsLegal(this Ship ship,
IEnumerable<Ship> ships,
Size board,
Point location,
ShipOrientation direction)
{
var temp = new Ship(ship.Length);
temp.Place(location, direction);
if (!temp.GetAllLocations().All(p => p.IsIn(board)))
return false;
return ships.Where(s => s.IsPlaced).All(s => !s.ConflictsWith(temp));
}
public static bool IsTouching(this Point a, Point b)
{
return (a.X == b.X - 1 || a.X == b.X + 1) &&
(a.Y == b.Y - 1 || a.Y == b.Y + 1);
}
public static bool IsTouching(this Ship ship,
IEnumerable<Ship> ships,
Point location,
ShipOrientation direction)
{
var temp = new Ship(ship.Length);
temp.Place(location, direction);
var occupied = new HashSet<Point>(ships
.Where(s => s.IsPlaced)
.SelectMany(s => s.GetAllLocations()));
if (temp.GetAllLocations().Any(p => occupied.Any(b => b.IsTouching(p))))
return true;
return false;
}
public static ReadOnlyCollection<Ship> MakeShips(params int[] lengths)
{
return new System.Collections.ObjectModel.ReadOnlyCollection<Ship>(
lengths.Select(l => new Ship(l)).ToList());
}
public static IEnumerable<T> Shuffle<T>(this IEnumerable<T> source, Rand rand)
{
T[] elements = source.ToArray();
// Note i > 0 to avoid final pointless iteration
for (int i = elements.Length - 1; i > 0; i--)
{
// Swap element "i" with a random earlier element it (or itself)
int swapIndex = rand.Next(i + 1);
T tmp = elements[i];
elements[i] = elements[swapIndex];
elements[swapIndex] = tmp;
}
// Lazily yield (avoiding aliasing issues etc)
foreach (T element in elements)
{
yield return element;
}
}
public static T RandomOrDefault<T>(this IEnumerable<T> things, Rand rand)
{
int count = things.Count();
if (count == 0)
return default(T);
return things.ElementAt(rand.Next(count));
}
}
}
Something I end up using a lot.
enum OpponentsBoardState
{
Unknown = 0,
Miss,
MustBeEmpty,
Hit,
}
Randomization. Secure but testable, useful for testing.
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Drawing;
namespace Battleship.ShuggyCoUk
{
public class Rand
{
Random r;
public Rand()
{
var rand = System.Security.Cryptography.RandomNumberGenerator.Create();
byte[] b = new byte[4];
rand.GetBytes(b);
r = new Random(BitConverter.ToInt32(b, 0));
}
public int Next(int maxValue)
{
return r.Next(maxValue);
}
public double NextDouble(double maxValue)
{
return r.NextDouble() * maxValue;
}
public T Pick<T>(IEnumerable<T> things)
{
return things.ElementAt(Next(things.Count()));
}
public T PickBias<T>(Func<T, double> bias, IEnumerable<T> things)
{
double d = NextDouble(things.Sum(x => bias(x)));
foreach (var x in things)
{
if (d < bias(x))
return x;
d -= bias(x);
}
throw new InvalidOperationException("fell off the end!");
}
}
}