这确实有点工作要找到它具有完美的策略给它一个机器人实施,不知道食物来源的位置和数量。
任何给定的机器人策略可能不会在每次运行中产生最大可能的收获。所以问题在于,在多次模拟运行中哪种策略最为成功。
要找到适合方形类型的给定的统计分布体面策略(P(快餐),P(slowFood),P(障碍)),人们可能会想出以下的想法:
让博特( npatch)是一个寻找npatch食物斑点的机器人。在第一次食物补丁找到第二次食物补丁之前,这种策略会消耗掉它发现的东西,依此类推。当它访问npatch食物来源(或没有发现更多食物补丁)时,它会返回到第一个找到并重新收获的食物来源。
这类僵尸工具(Bot(npatch))现在可以在统计相关数量的仿真运行中相互竞争。最好的机器人是比赛的胜利者。
这种方法可以被认为是基于遗传算法的启发,但没有混合任何基因,而只是迭代所有基因(1..npatch)。也许有人有一个想法如何把这个想法变成一个完全的遗传算法。这可能涉及转向Bot(npatch,searchStrategy),然后有多个基因来应用遗传算法。
每当模拟参数发生变化时,竞争必须重复,显然取决于世界上的食物补丁数量,如果某些食物补丁可能会或可能无法获得又一个食物补丁已知。
下面的代码是用F#编写的,它是这个问题的模拟器(如果我的要求正确,那就是......)。写一个新的机器人就像编写一个函数一样简单,然后传递给模拟器。
想想这个我的复活节彩蛋,适合那些想尝试自己的机器人的人。
我写的2个机器人被称为“marvinRobot”,它可以做Marvin会做的事情,而“lazyRobot”是一个在它找到的第一个食物来源营地的机器人。
type Square =
| Empty
| Obstacle
| Food of float * (float -> float) // available * growth
| Unknown
let rnd = new System.Random()
let grow p a =
let r = rnd.NextDouble()
if r < p then a + 1.0
else a
let slowGrowth a = grow 0.01 a
let fastGrowth a = grow 0.02 a
let eatPerTick = 1.0
let maxFoodPerSquare = 20.0
let randomPick values =
let count = List.length values
let r = rnd.Next(0,count-1)
values.Item(r)
type World = Square[,]
let randomSquare pobstacle pfood =
let r = rnd.NextDouble()
match r with
| x1 when x1 < pobstacle -> Obstacle
| x2 when x2 < (pobstacle + pfood) && x2 >= pobstacle ->
Food(rnd.NextDouble() * maxFoodPerSquare, randomPick [slowGrowth; fastGrowth])
| _ -> Empty
let createRandomWorld n pobstacle pfood =
Array2D.init n n (fun col row -> randomSquare pobstacle pfood)
let createUnknownWorld n =
Array2D.create n n Unknown
type Position = { Column : int; Row : int }
type RoboState = { Memory : Square[,]; Pos : Position; Heading : Position }
type RoboAction =
| TurnRight
| TurnLeft
| MoveOne
| Eat
| Idle
type RoboActor = World -> RoboState -> RoboAction
let right heading : Position =
match heading with
| { Column = 0; Row = 1 } -> { Column = -1; Row = 0 }
| { Column = -1; Row = 0 } -> { Column = 0; Row = -1 }
| { Column = 0; Row = -1 } -> { Column = 1; Row = 0 }
| { Column = 1; Row = 0 } -> { Column = 0; Row = 1 }
| _ -> failwith "Invalid heading!"
let left heading : Position =
match heading with
| { Column = -1; Row = 0 } -> { Column = 0; Row = 1 }
| { Column = 0; Row = -1 } -> { Column = -1; Row = 0 }
| { Column = 1; Row = 0 } -> { Column = 0; Row = -1 }
| { Column = 0; Row = 1 } -> { Column = 1; Row = 0 }
| _ -> failwith "Invalid heading!"
let checkAccess n position =
let inRange v = v >= 0 && v < n
(inRange position.Column) && (inRange position.Row)
let tickWorld world =
world
|> Array2D.map
(fun sq ->
match sq with
| Empty -> Empty
| Obstacle -> Obstacle
| Food(a,r) -> Food(min (r a) maxFoodPerSquare, r)
| Unknown -> Unknown
)
let rec step robot world roboState i imax acc =
if i < imax then
let action = robot world roboState
match action with
| TurnRight ->
let rs1 = { roboState with Heading = right roboState.Heading }
let wrld1 = tickWorld world
step robot wrld1 rs1 (i+1) imax acc
| TurnLeft ->
let rs1 = { roboState with Heading = left roboState.Heading }
let wrld1 = tickWorld world
step robot wrld1 rs1 (i+1) imax acc
| MoveOne ->
let rs1 =
let c =
{ Column = roboState.Pos.Column + roboState.Heading.Column
Row = roboState.Pos.Row + roboState.Heading.Row
}
if checkAccess (Array2D.length1 world) c
then
match world.[c.Column,c.Row] with
| Obstacle ->
roboState.Memory.[c.Column,c.Row] <- Obstacle
roboState
| _ -> { roboState with Pos = c }
else
roboState
let wrld1 = tickWorld world
step robot wrld1 rs1 (i+1) imax acc
| Eat ->
let eat,acc1 =
match world.[roboState.Pos.Column,roboState.Pos.Row] with
| Empty -> Empty,acc
| Obstacle -> Obstacle,acc
| Food(a,r) ->
let eaten = if a >= eatPerTick then eatPerTick else 0.0
printfn "eating %f carrots" eaten
Food(a - eaten, r),eaten + acc
| Unknown -> Unknown,acc
world.[roboState.Pos.Column,roboState.Pos.Row] <- eat
let wrld1 = tickWorld world
step robot wrld1 roboState (i+1) imax acc1
| Idle ->
step robot (tickWorld world) roboState (i+1) imax acc
else
acc
let initRoboState n =
{ Memory = createUnknownWorld n;
Pos = { Column = 0; Row = 0;};
Heading = {Column = 1; Row = 0}
}
let simulate n pobstacle pfood imax robot =
let w0 = createRandomWorld n pobstacle pfood
let r0 = initRoboState n
printfn "World: %A" w0
printfn "Initial Robo State: %A" r0
let result = step robot w0 r0 0 imax 0.0
printfn "Final Robo State: %A" r0
result
// Not that Marvin would care, but the rule for this simulator is that the
// bot may only inspect the square in the world at the current position.
// This means, IT CANNOT SEE the neighboring squares.
// This means, that if there is a obstacle next to current square,
// it costs a simulation tick to find out, trying to bump against it.
// Any access to other squares in world is considered cheating!
// world is passed in spite of all said above to allow for alternate rules.
let marvinRobot world roboState =
Idle
// Tries to find a square with food, then stays there, eating when there is something to eat.
let lazyRobot (world : World) (roboState : RoboState) =
let search() =
let status action : RoboAction =
match action with
| TurnLeft -> printfn "%A TurnLeft at %A (heading: %A)" world.[roboState.Pos.Column,roboState.Pos.Row] roboState.Pos roboState.Heading
| TurnRight -> printfn "%ATurnRight at %A (heading: %A)" world.[roboState.Pos.Column,roboState.Pos.Row] roboState.Pos roboState.Heading
| MoveOne -> printfn "%A MoveOne at %A (heading: %A)" world.[roboState.Pos.Column,roboState.Pos.Row] roboState.Pos roboState.Heading
| Idle -> printfn "%A Idle at %A (heading: %A)" world.[roboState.Pos.Column,roboState.Pos.Row] roboState.Pos roboState.Heading
| Eat -> printfn "%A Eat at %A (heading: %A)" world.[roboState.Pos.Column,roboState.Pos.Row] roboState.Pos roboState.Heading
action
let neighbors =
[ roboState.Heading, MoveOne;
(roboState.Heading |> right),TurnRight;
(roboState.Heading |> left),TurnLeft;
(roboState.Heading |> right |> right),TurnRight
]
|> List.map (fun (p,a) -> (p.Column,p.Row),a)
|> List.map (fun ((c,r),a) -> (roboState.Pos.Column + c,roboState.Pos.Row + r),a)
|> List.filter (fun ((c,r),a) -> checkAccess (Array2D.length1 world){Position.Column = c; Row = r})
|> List.sortBy (fun ((c,r),a) -> match roboState.Memory.[c,r] with | Food(_,_) -> 0 | Unknown -> 1 | Empty -> 2 | Obstacle -> 3)
|> List.map (fun ((c,r),a) -> { Column = c; Row = r},a)
if neighbors.IsEmpty then failwith "It's a trap!" // can happen if bot is surrounded by obstacles, e.g. in a corner
else
let p,a = neighbors.Head
status a
roboState.Memory.[roboState.Pos.Column, roboState.Pos.Row] <-
world.[roboState.Pos.Column,roboState.Pos.Row]
match world.[roboState.Pos.Column,roboState.Pos.Row] with
| Food(a,_) ->
printfn "Found food at %A" roboState.Pos
Eat
| _ ->
search()
//simulate 10 0.1 0.05 2000 marvinRobot
simulate 10 0.1 0.1 2000 lazyRobot
最后不是至少提示:如果您0.0食物补丁模拟,你的机器人应该参观了地图上所有的广场。如果它没有这样做,它肯定不是一个好的机器人;)
(1)事先知道胡萝卜的位置? (2)事先知道障碍物吗? (3)即使有了这些假设,问题仍然很难 - 因为它很容易从哈密尔顿路径/旅行销售人员问题中减少。在没有这些假设的情况下获得最佳解决方案将变得更加困难。 – amit 2015-04-05 15:41:29
(1)不知道胡萝卜的位置是未知的 (2)障碍物也没有 – 2015-04-05 15:42:11
已经收获的胡萝卜点会发生什么?他们在一段时间后会重新长大吗? – BitTickler 2015-04-05 15:42:45