游戏中的 Q-learning 未按预期工作

Question

我试图在我编写的一个简单游戏中实施 Q 学习。该游戏基于玩家必须 "jump" 来避开迎面而来的盒子。

我设计的系统有两个动作； jump和do_nothing和states是到下一个block的距离（分和底，保证不会出现大量的states）。

我的问题似乎是我的算法实现没有考虑 "future reward"，所以它最终跳错了时间。

这是我对 Q-learning 算法的实现；

JumpGameAIClass.prototype.getQ = function getQ(state) {
    if (!this.Q.hasOwnProperty(state)) {
        this.Q[state] = {};

        for (var actionIndex = 0; actionIndex < this.actions.length; actionIndex++) {
            var action = this.actions[actionIndex];

            this.Q[state][action] = 0;
        }
    }

    return this.Q[state];
};

JumpGameAIClass.prototype.getBlockDistance = function getBlockDistance() {
    var closest = -1;

    for (var blockIndex = 0; blockIndex < this.blocks.length; blockIndex++) {
        var block = this.blocks[blockIndex];

        var distance = block.x - this.playerX;

        if (distance >= 0 && (closest === -1 || distance < closest)) {
            closest = distance;
        }
    }

    return Math.max(0, Math.floor(closest * this.resolution));
};

JumpGameAIClass.prototype.getActionWithHighestQ = function getActionWithHighestQ(distance) {
    var jumpReward = this.getQ(distance)[this.actions[0]];
    var doNothingReward = this.getQ(distance)[this.actions[1]];

    if (jumpReward > doNothingReward) {
        return this.actions[0];
    } else if (doNothingReward > jumpReward) {
        return this.actions[1];
    } else {
        if (!this.canJump()) {
            return this.actions[1];
        }

        return this.actions[Math.floor(Math.random() * this.actions.length)];
    }
};

JumpGameAIClass.prototype.getActionEpsilonGreedy = function getActionEpsilonGreedy() {
    // We can't jump while in mid-air
    if (!this.canJump()) {
        return this.actions[1];
    }

    if (Math.random() < this.epsilon) {
        return this.actions[Math.floor(Math.random() * this.actions.length)];
    } else {
        return this.getActionWithHighestQ(this.getBlockDistance());
    }
};

JumpGameAIClass.prototype.think = function think() {
    var reward = this.liveReward;

    if (this.score !== this.lastScore) {
        this.lastScore = this.score;
        reward = this.scoreReward;
    } else if (!this.playerAlive) {
        reward = this.deathReward;
    }

    this.drawDistance();

    var distance = this.getBlockDistance(),
        maxQ = this.getQ(distance)[this.getActionWithHighestQ(distance)],
        previousQ = this.getQ(this.lastDistance)[this.lastAction];

    this.getQ(this.lastDistance)[this.lastAction] = previousQ + this.alpha * (reward + (this.gamma * maxQ) - previousQ);

    this.lastAction = this.getActionEpsilonGreedy();
    this.lastDistance = distance;

    switch (this.lastAction) {
        case this.actions[0]:
            this.jump();
            break;
    }
};

下面是它使用的一些属性：

epsilon: 0.05,
alpha: 1,
gamma: 1,
resolution: 0.1,
actions: [ 'jump', 'do_nothing' ],
Q: {},
liveReward: 0,
scoreReward: 100,
deathReward: -1000,
lastAction: 'do_nothing',
lastDistance: 0,
lastScore: 0

我不得不使用 lastAction/lastDistance 来计算 Q，因为我不能使用当前数据（将作用于之前帧中执行的动作）。

在所有渲染和游戏内容（物理、控制、死亡等）完成后，每帧调用一次 think 方法。

var JumpGameAIClass = function JumpGame(canvas) {
    Game.JumpGame.call(this, canvas);

    Object.defineProperties(this, {
        epsilon: {
            value: 0.05
        },

        alpha: {
            value: 1
        },

        gamma: {
            value: 1
        },

        resolution: {
            value: 0.1
        },

        actions: {
            value: [ 'jump', 'do_nothing' ]
        },

        Q: {
            value: { },
            writable: true
        },

        liveReward: {
            value: 0
        },

        scoreReward: {
            value: 100
        },

        deathReward: {
            value: -1000
        },

        lastAction: {
            value: 'do_nothing',
            writable: true
        },

        lastDistance: {
            value: 0,
            writable: true
        },

        lastScore: {
            value: 0,
            writable: true
        }
    });
};

JumpGameAIClass.prototype = Object.create(Game.JumpGame.prototype);

JumpGameAIClass.prototype.getQ = function getQ(state) {
    if (!this.Q.hasOwnProperty(state)) {
        this.Q[state] = {};

        for (var actionIndex = 0; actionIndex < this.actions.length; actionIndex++) {
            var action = this.actions[actionIndex];

            this.Q[state][action] = 0;
        }
    }

    return this.Q[state];
};

JumpGameAIClass.prototype.getBlockDistance = function getBlockDistance() {
    var closest = -1;

    for (var blockIndex = 0; blockIndex < this.blocks.length; blockIndex++) {
        var block = this.blocks[blockIndex];

        var distance = block.x - this.playerX;

        if (distance >= 0 && (closest === -1 || distance < closest)) {
            closest = distance;
        }
    }

    return Math.max(0, Math.floor(closest * this.resolution));
};

JumpGameAIClass.prototype.getActionWithHighestQ = function getActionWithHighestQ(distance) {
    var jumpReward = this.getQ(distance)[this.actions[0]];
    var doNothingReward = this.getQ(distance)[this.actions[1]];

    if (jumpReward > doNothingReward) {
        return this.actions[0];
    } else if (doNothingReward > jumpReward) {
        return this.actions[1];
    } else {
        if (!this.canJump()) {
            return this.actions[1];
        }

        return this.actions[Math.floor(Math.random() * this.actions.length)];
    }
};

JumpGameAIClass.prototype.getActionEpsilonGreedy = function getActionEpsilonGreedy() {
    if (!this.canJump()) {
        return this.actions[1];
    }

    if (Math.random() < this.epsilon) {
        return this.actions[Math.floor(Math.random() * this.actions.length)];
    } else {
        return this.getActionWithHighestQ(this.getBlockDistance());
    }
};

JumpGameAIClass.prototype.onDeath = function onDeath() {
    this.restart();
};

JumpGameAIClass.prototype.think = function think() {
    var reward = this.liveReward;

    if (this.score !== this.lastScore) {
        this.lastScore = this.score;
        reward = this.scoreReward;
    } else if (!this.playerAlive) {
        reward = this.deathReward;
    }

    this.drawDistance();

    var distance = this.getBlockDistance(),
        maxQ = this.getQ(distance)[this.getActionWithHighestQ(distance)],
        previousQ = this.getQ(this.lastDistance)[this.lastAction];

    this.getQ(this.lastDistance)[this.lastAction] = previousQ + this.alpha * (reward + (this.gamma * maxQ) - previousQ);

    this.lastAction = this.getActionEpsilonGreedy();
    this.lastDistance = distance;

    switch (this.lastAction) {
        case this.actions[0]:
            this.jump();
            break;
    }
};

JumpGameAIClass.prototype.drawDistance = function drawDistance() {
    this.context.save();

    this.context.textAlign = 'center';
    this.context.textBaseline = 'bottom';

    this.context.fillText('Distance: ' + this.getBlockDistance(), this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.textBaseline = 'top';

    this.context.fillText('Last Distance: ' + this.lastDistance, this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.restore();
};

JumpGameAIClass.prototype.onFrame = function onFrame() {
    Game.JumpGame.prototype.onFrame.apply(this, arguments);

    this.think();
}

Game.JumpGameAI = JumpGameAIClass;

body {
    background-color: #EEEEEE;
    text-align: center;
}

canvas#game {
    background-color: #FFFFFF;
    border: 1px solid #DDDDDD;
}

<!DOCTYPE HTML>
<html lang="en">
<head>
    <title>jump</title>
</head>
<body>
    <canvas id="game" width="512" height="512">
        <h1>Your browser doesn't support canvas!</h1>
    </canvas>
  
    <script src="https://raw.githubusercontent.com/cagosta/requestAnimationFrame/master/app/requestAnimationFrame.js"></script>
  
    <!-- https://gist.github.com/jackwilsdon/d06bffa6b32c53321478 -->
  
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/4e467f82590e76543bf55ff788504e26afc3d694/game.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2b7ce2c3dd268c4aef9ad27316edb0b235ad0d06/canvasgame.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2696c72e001e48359a6ce880f1c475613fe359f5/jump.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/249c92f3385757b6edf2ceb49e26f14b89ffdcfe/bootstrap.js"></script>
</body>

Answer 1

你基本上有简化版的:

来源：Flappy Bird RL

我使用的值：

    epsilon: {
        value: 0.01
    },
    alpha: {
        value: 0.7
    },
    gamma: {
        value: 0.9
    },
    resolution: {
        value: 0.1
    },  
    liveReward: {
        value: 10
    },
    scoreReward: {
        value: -100
    },
    deathReward: {
        value: 1000
    },

它在前 20 次尝试中轻松超过 100。

Q-learning 可以用时间逻辑来描述

Q(s, a)=r(s,a)+gamma*max_a'(Q(s', a'))

哪里

r(s,a) = r = 立即奖励
gamma = 延迟奖励与立即奖励的相对值（0 到 1）
s' = 动作后的新状态a
a = 状态 s
a' = 状态 s'

你应该执行它

Select一个动作a并执行

对于每个状态-动作对 (s, a)，将 table 条目 Q(s, a) 初始化为零
观察当前状态s
永远做：
- Select一个动作a并执行
- 立即获得奖励 r 即 Q(s, a)
- 观察新状态s'
- 更新 table 条目 Q(s,a)=r(s,a)+伽玛*max_a'(Q(s',a'))
- s=s'

Answer 2

你的算法实现没问题，只是需要调整一些参数。

如果您为生活分配一些奖励，在我的示例中为 10，并将 epsilon 设置为 0，您将获得一个获胜的 AI。

示例：

var JumpGameAIClass = function JumpGame(canvas) {
    Game.JumpGame.call(this, canvas);

    Object.defineProperties(this, {
        epsilon: {
            value: 0
        },

        alpha: {
            value: 1
        },

        gamma: {
            value: 1
        },

        resolution: {
            value: 0.1
        },

        actions: {
            value: [ 'jump', 'do_nothing' ]
        },

        Q: {
            value: { },
            writable: true
        },

        liveReward: {
            value: 0
        },

        scoreReward: {
            value: 100
        },

        deathReward: {
            value: -1000
        },

        lastAction: {
            value: 'do_nothing',
            writable: true
        },

        lastDistance: {
            value: 0,
            writable: true
        },

        lastScore: {
            value: 0,
            writable: true
        }
    });
};

JumpGameAIClass.prototype = Object.create(Game.JumpGame.prototype);

JumpGameAIClass.prototype.getQ = function getQ(state) {
    if (!this.Q.hasOwnProperty(state)) {
        this.Q[state] = {};

        for (var actionIndex = 0; actionIndex < this.actions.length; actionIndex++) {
            var action = this.actions[actionIndex];

            this.Q[state][action] = 0;
        }
    }

    return this.Q[state];
};

JumpGameAIClass.prototype.getBlockDistance = function getBlockDistance() {
    var closest = -1;

    for (var blockIndex = 0; blockIndex < this.blocks.length; blockIndex++) {
        var block = this.blocks[blockIndex];

        var distance = block.x - this.playerX;

        if (distance >= 0 && (closest === -1 || distance < closest)) {
            closest = distance;
        }
    }

    return Math.max(0, Math.floor(closest * this.resolution));
};

JumpGameAIClass.prototype.getActionWithHighestQ = function getActionWithHighestQ(distance) {
    var jumpReward = this.getQ(distance)[this.actions[0]];
    var doNothingReward = this.getQ(distance)[this.actions[1]];
    
    if (!this.canJump()) {
        return this.actions[1];
    } else if (jumpReward > doNothingReward) {
        return this.actions[0];
    } else if (doNothingReward > jumpReward) {
        return this.actions[1];
    } else {   
        return this.actions[Math.floor(Math.random() * this.actions.length)];
    }
};

JumpGameAIClass.prototype.getActionEpsilonGreedy = function getActionEpsilonGreedy() {
    if (!this.canJump()) {
        return this.actions[1];
    }

    if (Math.random() < this.epsilon) {
        return this.actions[Math.floor(Math.random() * this.actions.length)];
    } else {
        return this.getActionWithHighestQ(this.getBlockDistance());
    }
};

JumpGameAIClass.prototype.onDeath = function onDeath() {
    this.restart();
};

JumpGameAIClass.prototype.think = function think() {
    var reward = this.liveReward;

    if (this.score !== this.lastScore) {
        this.lastScore = this.score;
        reward = this.scoreReward;
    } else if (!this.playerAlive) {
        reward = this.deathReward;
    }

    this.drawDistance();

    var distance = this.getBlockDistance(),
        maxQ = this.playerAlive ? this.getQ(distance)[this.getActionWithHighestQ(distance)] : 0,
        previousQ = this.getQ(this.lastDistance)[this.lastAction];

    this.getQ(this.lastDistance)[this.lastAction] = previousQ + this.alpha * (reward + (this.gamma * maxQ) - previousQ);

    this.lastAction = this.getActionEpsilonGreedy();
    this.lastDistance = distance;

    switch (this.lastAction) {
        case this.actions[0]:
            this.jump();
            break;
    }
};

JumpGameAIClass.prototype.drawDistance = function drawDistance() {
    this.context.save();

    this.context.textAlign = 'center';
    this.context.textBaseline = 'bottom';

    this.context.fillText('Distance: ' + this.getBlockDistance(), this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.textBaseline = 'top';

    this.context.fillText('Last Distance: ' + this.lastDistance, this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.restore();
};

JumpGameAIClass.prototype.onFrame = function onFrame() {
    Game.JumpGame.prototype.onFrame.apply(this, arguments);

    this.think();
}

Game.JumpGameAI = JumpGameAIClass;

body {
    background-color: #EEEEEE;
    text-align: center;
}

canvas#game {
    background-color: #FFFFFF;
    border: 1px solid #DDDDDD;
}

<!DOCTYPE HTML>
<html lang="en">
<head>
    <title>jump</title>
</head>
<body>
    <canvas id="game" width="512" height="512">
        <h1>Your browser doesn't support canvas!</h1>
    </canvas>
  
    <script src="https://raw.githubusercontent.com/cagosta/requestAnimationFrame/master/app/requestAnimationFrame.js"></script>
  
    <!-- https://gist.github.com/jackwilsdon/d06bffa6b32c53321478 -->
  
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/4e467f82590e76543bf55ff788504e26afc3d694/game.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2b7ce2c3dd268c4aef9ad27316edb0b235ad0d06/canvasgame.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2696c72e001e48359a6ce880f1c475613fe359f5/jump.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/249c92f3385757b6edf2ceb49e26f14b89ffdcfe/bootstrap.js"></script>
</body>

更新：

对此进行了更多思考，虽然我的示例似乎有效，但它并不正确。

令人高兴的是，因为跳跃的结果直到未来的多次迭代才知道，所以为生活分配即时奖励将导致重复在每个状态下首先做出的随机决定，直到决策的最终结果会传回各州。

根据游戏的物理原理，玩家的跳跃距离小于方块间距，这意味着清除方块的跳跃将比最后一个方块的起飞点离下一个方块更远，因此可以再次进行相同的跳跃。因此，如果在第一个块之前进行 "good" 跳转，系统将立即收敛到一个成功的模式。如果游戏的物理机制不同或进行了 "bad" 跳跃，则此 AI 可能无法自行纠正。

问题是系统的状态实际上有两个部分，blockDistance 和 playerY。如果决策中不包含 playerY 状态，跳跃的结果将无法正确传播回其原点。

在这个简单的游戏中，您可以通过偏向于不采取任何行动的决定来解决这个问题。由于仅基于距离的决策状态是完整的，前提是您不跳跃，因此不跳跃的结果（即死亡）将正确地传播回每个距离的不跳跃决策。它仍然有点古怪，因为一旦你跳跃，奖励的传播将不正确，但你现在可以看到它学习都是一样的。

示例：

var JumpGameAIClass = function JumpGame(canvas) {
    Game.JumpGame.call(this, canvas);

    Object.defineProperties(this, {
        epsilon: {
            value: 0
        },

        alpha: {
            value: 1
        },

        gamma: {
            value: 1
        },

        resolution: {
            value: 0.1
        },

        actions: {
            value: [ 'jump', 'do_nothing' ]
        },

        Q: {
            value: { },
            writable: true
        },

        liveReward: {
            value: 10
        },

        scoreReward: {
            value: 100
        },

        deathReward: {
            value: -1000
        },

        lastAction: {
            value: 'do_nothing',
            writable: true
        },

        lastDistance: {
            value: 0,
            writable: true
        },

        lastScore: {
            value: 0,
            writable: true
        }
    });
};

JumpGameAIClass.prototype = Object.create(Game.JumpGame.prototype);

JumpGameAIClass.prototype.getQ = function getQ(state) {
    if (!this.Q.hasOwnProperty(state)) {
        this.Q[state] = {};

        for (var actionIndex = 0; actionIndex < this.actions.length; actionIndex++) {
            var action = this.actions[actionIndex];

            this.Q[state][action] = 0;
        }
    }

    return this.Q[state];
};

JumpGameAIClass.prototype.getBlockDistance = function getBlockDistance() {
    var closest = -1;

    for (var blockIndex = 0; blockIndex < this.blocks.length; blockIndex++) {
        var block = this.blocks[blockIndex];

        var distance = block.x - this.playerX;

        if (distance >= 0 && (closest === -1 || distance < closest)) {
            closest = distance;
        }
    }

    return Math.max(0, Math.floor(closest * this.resolution));
};

JumpGameAIClass.prototype.getActionWithHighestQ = function getActionWithHighestQ(distance) {
    var jumpReward = this.getQ(distance)[this.actions[0]];
    var doNothingReward = this.getQ(distance)[this.actions[1]];

if (!this.canJump() || doNothingReward >= jumpReward) {
 return this.actions[1];
} else {
 return this.actions[0];
}    
};

JumpGameAIClass.prototype.getActionEpsilonGreedy = function getActionEpsilonGreedy() {
    if (!this.canJump()) {
        return this.actions[1];
    }

    if (Math.random() < this.epsilon) {
        return this.actions[Math.floor(Math.random() * this.actions.length)];
    } else {
        return this.getActionWithHighestQ(this.getBlockDistance());
    }
};

JumpGameAIClass.prototype.onDeath = function onDeath() {
    this.restart();
};

JumpGameAIClass.prototype.think = function think() {
    var reward = this.liveReward;

    if (this.score !== this.lastScore) {
        this.lastScore = this.score;
        reward = this.scoreReward;
    } else if (!this.playerAlive) {
        reward = this.deathReward;
    }

    this.drawDistance();

    var distance = this.getBlockDistance(),
        maxQ = this.playerAlive ? this.getQ(distance)[this.getActionWithHighestQ(distance)] : 0,
        previousQ = this.getQ(this.lastDistance)[this.lastAction];

    this.getQ(this.lastDistance)[this.lastAction] = previousQ + this.alpha * (reward + (this.gamma * maxQ) - previousQ);

    this.lastAction = this.getActionEpsilonGreedy();
    this.lastDistance = distance;

    switch (this.lastAction) {
        case this.actions[0]:
            this.jump();
            break;
    }
};

JumpGameAIClass.prototype.drawDistance = function drawDistance() {
    this.context.save();

    this.context.textAlign = 'center';
    this.context.textBaseline = 'bottom';

    this.context.fillText('Distance: ' + this.getBlockDistance(), this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.textBaseline = 'top';

    this.context.fillText('Last Distance: ' + this.lastDistance, this.canvasWidth / 2, this.canvasHeight / 4);

    this.context.restore();
};

JumpGameAIClass.prototype.onFrame = function onFrame() {
    Game.JumpGame.prototype.onFrame.apply(this, arguments);

    this.think();
}

Game.JumpGameAI = JumpGameAIClass;

body {
    background-color: #EEEEEE;
    text-align: center;
}

canvas#game {
    background-color: #FFFFFF;
    border: 1px solid #DDDDDD;
}

<!DOCTYPE HTML>
<html lang="en">
<head>
    <title>jump</title>
</head>
<body>
    <canvas id="game" width="512" height="512">
        <h1>Your browser doesn't support canvas!</h1>
    </canvas>
  
    <script src="https://raw.githubusercontent.com/cagosta/requestAnimationFrame/master/app/requestAnimationFrame.js"></script>
  
    <!-- https://gist.github.com/jackwilsdon/d06bffa6b32c53321478 -->
  
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/4e467f82590e76543bf55ff788504e26afc3d694/game.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2b7ce2c3dd268c4aef9ad27316edb0b235ad0d06/canvasgame.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/2696c72e001e48359a6ce880f1c475613fe359f5/jump.js"></script>
    <script src="https://cdn.rawgit.com/jackwilsdon/d06bffa6b32c53321478/raw/249c92f3385757b6edf2ceb49e26f14b89ffdcfe/bootstrap.js"></script>
</body>

游戏中的 Q-learning 未按预期工作

Q-learning in game not working as expected

artificial-intelligence

game-ai

q-learning