战舰!

早在2003年(当时我17岁),我参加了一场《战舰》AI编码比赛。尽管我输了那场比赛,但我从中获得了很多乐趣,也学到了很多东西。

现在,我想恢复这个比赛,在搜索最好的战舰AI。

下面是这个框架,现在托管在Bitbucket上。

获胜者将获得+450声望奖励!比赛将于2009年11月17日开始。17号零时之前的投稿和编辑将不被接受。(中央标准时间) 尽早提交你的作品,这样你就不会错过机会!

为了保持这个目标,请遵循比赛的精神。

游戏规则:

游戏在10x10的网格上进行。 每个参赛者将5艘船(长度为2、3、3、4、5)中的每一艘放在他们的网格上。 没有船只可以重叠,但它们可以相邻。 然后选手们轮流向对手射击。 游戏的一个变体允许每次齐射多次,每艘幸存的船一次。 如果击球沉、命中或未命中,对手将通知选手。 当任何一名玩家的所有船只都沉没时,游戏就结束了。

比赛规则:

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.

得分:

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.

好运!玩得开心!


编辑1: 多亏弗里德,他在飞船上发现了一个错误。是否是可用的函数。问题已经解决了。请下载框架的更新版本。

编辑2: 由于人们对将统计数据持久化到磁盘等非常感兴趣,所以我添加了一些非计时设置和删除事件,它们应该能够提供所需的功能。这是一个半突破性的变化。也就是说:修改了接口,添加了功能,但不需要body。请下载框架的更新版本。

编辑3: 错误修复1:GameWon和GameLost只在超时的情况下被调用。 错误修复2:如果引擎在每一款游戏中都暂停计时,那么竞争将永远不会结束。 请下载框架的更新版本。

编辑4: 比赛结果:


当前回答

我总是喜欢从中间开始,从那个点螺旋离开,在任何其他点之间留下不超过1个空白,以解释那个该死的子…射击间隔取决于哪艘船被击沉。如果b舰是最后一艘,射击之间只需要留出4个空间来减少浪费

其他回答

如果你强迫自己进行分析,那么你可能会发现随机对手的机制非常低效。它允许自己重新选择已经目标的位置,并让框架强制它重复,直到它击中它还没有触及的位置,或者每次移动的时间限制到期。

这个对手有类似的行为(有效的位置分布是相同的),它只是自己进行完整性检查,每次调用只消耗一个随机数生成(平摊)。

这使用了我的扩展/库答案中的类,我只提供关键方法/状态。

Shuffle源自Jon Skeet的回答

class WellBehavedRandomOpponent : IBattleShipOpponent
{
    Rand rand = new Rand();
    List<Point> guesses;
    int nextGuess = 0;

    public void PlaceShips(IEnumerable<Ship> ships)
    {
        BoardView<bool> board = new BoardView<bool>(BoardSize);
        var AllOrientations = new[] {
            ShipOrientation.Horizontal,
            ShipOrientation.Vertical };

        foreach (var ship in ships)
        {
            while (!ship.IsPlaced)
            {
                var l = rand.Pick(board.Select(c => c.Location));
                var o = rand.Pick(AllOrientations);
                if (ship.IsLegal(ships, BoardSize, l, o))
                    ship.Place(l, o);
            }
        }
    }

    public void NewGame(Size size, TimeSpan timeSpan)
    {
        var board = new BoardView<bool>(size);
        this.guesses = new List<Point>(
            board.Select(x => x.Location).Shuffle(rand));
        nextGuess = 0;
    }

    public System.Drawing.Point GetShot()
    {
        return guesses[nextGuess++];
    }

    // empty methods left out 
}

关于竞争引擎的一些评论:

NewGame参数:

如果IBattleshipOpponent::NewGame用于游戏前设置,并接受一个boardsize,那么它还应该接受一个船的列表及其各自的大小。在不考虑船舶配置变化的情况下,允许可变的船板尺寸是没有意义的。

船舶是密封的:

我看不出为什么船班是封闭的。除了其他基本的东西,我想船有一个名字,所以我可以输出消息(“你沉没了我的{0}”,船.Name);我还考虑了其他扩展,所以我认为Ship应该是可继承的。

时间限制:

虽然1秒的时间限制对比赛规则来说是有意义的,但它完全扰乱了调试。BattleshipCompetition应该有一个容易忽略时间冲突的设置,以帮助开发/调试。我还建议调查System.Diagnostics。Process::UserProcessorTime / Privileged ProcessorTime / TotalProcessorTime用于更准确地查看正在使用的时间。

沉没的船只:

当你击沉对手的船时,当前的API会通知你:

ShotHit(Point shot, bool sunk);

但不是你击沉了哪艘船!我认为这是人类战列舰规则的一部分,你必须宣布“你击沉了我的战列舰!”(或驱逐舰,或潜艇等)。

当AI试图清除相互碰撞的船只时,这一点尤其重要。我想要求API更改为:

ShotHit(Point shot, Ship ship);

如果ship为非空,则意味着该射击是一个击沉射击,并且您知道您击沉了哪艘船以及它的长度。如果射击是一个非下沉射击,那么船是空的,你没有进一步的信息。

这是我的入口!(最天真的解决方案)

“随机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() { }
    }
}

我赞成每场比赛多打几场比赛。制作50款游戏只是抛硬币。我需要做1000个游戏,才能在测试算法之间找到合理的区别。

下载无畏1.2。

策略:

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.

萨里大学的詹姆斯·希瑟博士代表英国计算机学会举办了一次类似的比赛。

对资源的限制——即每回合的最大处理器时间,移动之间不能存储状态,最大堆大小。为了限制时间,AI可以在时间段内的任何时间点提交移动,并在回合结束时被要求移动。

非常有趣,更多信息请访问:http://www.bcsstudentcontest.com/

也许能给你更多的建议。