math-skill 1.0.0 → 2.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.en-US.md +313 -0
- package/README.md +313 -278
- package/agents/math-critic.en.md +235 -0
- package/agents/math-critic.md +237 -203
- package/commands/abstraction.md +11 -34
- package/commands/algorithmic-thinking.md +11 -34
- package/commands/ask.md +18 -21
- package/commands/axiomatization.md +11 -34
- package/commands/causal-inference.md +11 -34
- package/commands/discrete-combinatorial.md +11 -34
- package/commands/game-theory.md +11 -34
- package/commands/induction-analogy.md +11 -34
- package/commands/information-theory.md +11 -34
- package/commands/logic-deduction.md +11 -34
- package/commands/modeling.md +11 -37
- package/commands/optimization.md +11 -33
- package/commands/probability-statistics.md +11 -36
- package/commands/symmetry-invariance.md +11 -34
- package/commands/topological-thinking.md +11 -33
- package/commands/transformation.md +11 -33
- package/knowledge-base/overview.en.md +228 -0
- package/knowledge-base/overview.md +230 -230
- package/package.json +73 -59
- package/references/agentic-workflow.en.md +53 -0
- package/references/agentic-workflow.md +55 -0
- package/references/books/abstract-algebra.md +124 -0
- package/references/books/algebraic-geometry-rising-sea.md +171 -0
- package/references/books/differential-geometry.md +140 -0
- package/references/books/matrix-analysis.md +146 -0
- package/references/books/micro-lie-theory.md +116 -0
- package/references/books/optimization-ml.md +164 -0
- package/references/books/smooth-manifolds.md +105 -0
- package/references/gpu-friendly-math.en.md +65 -0
- package/references/gpu-friendly-math.md +67 -0
- package/references/inspiration.en.md +113 -0
- package/{docs → references}/inspiration.md +2 -0
- package/skills/abstraction/SKILL.en.md +117 -0
- package/skills/abstraction/SKILL.md +121 -264
- package/skills/abstraction/original-texts.en.md +163 -0
- package/skills/algorithmic-thinking/SKILL.en.md +132 -0
- package/skills/algorithmic-thinking/SKILL.md +138 -371
- package/skills/algorithmic-thinking/original-texts.en.md +253 -0
- package/skills/axiomatization/SKILL.en.md +144 -0
- package/skills/axiomatization/SKILL.md +151 -213
- package/skills/axiomatization/original-texts.en.md +154 -0
- package/skills/causal-inference/SKILL.en.md +147 -0
- package/skills/causal-inference/SKILL.md +151 -374
- package/skills/causal-inference/original-texts.en.md +136 -0
- package/skills/discrete-combinatorial/SKILL.en.md +124 -0
- package/skills/discrete-combinatorial/SKILL.md +131 -286
- package/skills/discrete-combinatorial/original-texts.en.md +184 -0
- package/skills/game-theory/SKILL.en.md +117 -0
- package/skills/game-theory/SKILL.md +123 -318
- package/skills/game-theory/original-texts.en.md +131 -0
- package/skills/induction-analogy/SKILL.en.md +145 -0
- package/skills/induction-analogy/SKILL.md +152 -310
- package/skills/induction-analogy/original-texts.en.md +140 -0
- package/skills/information-theory/SKILL.en.md +134 -0
- package/skills/information-theory/SKILL.md +140 -242
- package/skills/information-theory/original-texts.en.md +127 -0
- package/skills/logic-deduction/SKILL.en.md +130 -0
- package/skills/logic-deduction/SKILL.md +135 -280
- package/skills/logic-deduction/original-texts.en.md +160 -0
- package/skills/math-research-activator/SKILL.en.md +132 -0
- package/skills/math-research-activator/SKILL.md +136 -0
- package/skills/math-research-activator/original-texts.en.md +105 -0
- package/skills/{meta-selector → math-research-activator}/original-texts.md +104 -104
- package/skills/modeling/SKILL.en.md +135 -0
- package/skills/modeling/SKILL.md +139 -318
- package/skills/modeling/original-texts.en.md +162 -0
- package/skills/optimization/SKILL.en.md +129 -0
- package/skills/optimization/SKILL.md +135 -292
- package/skills/optimization/original-texts.en.md +167 -0
- package/skills/probability-statistics/SKILL.en.md +146 -0
- package/skills/probability-statistics/SKILL.md +151 -312
- package/skills/probability-statistics/original-texts.en.md +191 -0
- package/skills/symmetry-invariance/SKILL.en.md +135 -0
- package/skills/symmetry-invariance/SKILL.md +139 -358
- package/skills/symmetry-invariance/original-texts.en.md +206 -0
- package/skills/topological-thinking/SKILL.en.md +124 -0
- package/skills/topological-thinking/SKILL.md +128 -273
- package/skills/topological-thinking/original-texts.en.md +134 -0
- package/skills/transformation/SKILL.en.md +120 -0
- package/skills/transformation/SKILL.md +124 -264
- package/skills/transformation/original-texts.en.md +204 -0
- package/docs/CLAUDE.md +0 -187
- package/skills/meta-selector/SKILL.md +0 -188
|
@@ -0,0 +1,117 @@
|
|
|
1
|
+
---
|
|
2
|
+
name: game-theory
|
|
3
|
+
description: |
|
|
4
|
+
Trigger when optimal strategy depends on others' choices, needing Nash equilibrium computation, zero-sum/non-zero-sum analysis, mechanism design; or designing strategies for multi-agent systems, adversarial training, routing games.
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
# 🎯 Game Theory
|
|
8
|
+
|
|
9
|
+
> "Your optimal choice depends on others' choices — thinking must be interactive, not just deep."
|
|
10
|
+
>
|
|
11
|
+
> — Game Theory, Decision Theory, Mechanism Design
|
|
12
|
+
|
|
13
|
+
## Core Principle
|
|
14
|
+
|
|
15
|
+
**Strategic interaction — your best action depends on what others do, and theirs depend on what you do. Nash equilibrium: strategy profile where no player benefits from unilateral deviation.**
|
|
16
|
+
|
|
17
|
+
Key concepts: **Payoff Matrix**, **Dominant Strategy**, **Mixed Strategy**, **Pareto Optimality**. Classification dimensions: cooperative vs. non-cooperative, zero-sum vs. general-sum, simultaneous vs. sequential.
|
|
18
|
+
|
|
19
|
+
> **Mathematical Formalization**
|
|
20
|
+
>
|
|
21
|
+
> Formal definition: $n$ players, each with strategy set $S_i$ and payoff function $u_i: S_1 \times \cdots \times S_n \to \mathbb{R}$.
|
|
22
|
+
>
|
|
23
|
+
> Nash equilibrium: strategy profile $s^* = (s_1^*, \ldots, s_n^*)$ satisfying $u_i(s^*) \geq u_i(s_i, s^*_{-i})$ for all $s_i \in S_i$ and all $i$, where $s^*_{-i}$ denotes the strategies of all players except $i$. Mixed-strategy Nash equilibrium: probability distribution $\sigma^*$ such that $u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$.
|
|
24
|
+
>
|
|
25
|
+
> Dominant strategy: strict $u_i(s_i^*, s_{-i}) > u_i(s_i, s_{-i})$; weak dominance $\geq$ with at least one strict inequality.
|
|
26
|
+
>
|
|
27
|
+
> Zero-sum minimax: $\max_x \min_y f(x,y) = \min_y \max_x f(x,y)$.
|
|
28
|
+
|
|
29
|
+
## GPU-Friendliness (Cross-Cutting Check)
|
|
30
|
+
|
|
31
|
+
Game-solving maps to GPU with widely varying capability; pass through the eight-dimensional gate in `../../references/gpu-friendly-math.md`:
|
|
32
|
+
|
|
33
|
+
- **Matrix games / two-player zero-sum (minimax)**: Equivalent to linear programming (LP); payoff matrix operations are GEMM-compatible and batch-parallelizable — **friendly**.
|
|
34
|
+
- **Mixed-strategy equilibrium computation**: Linear systems / indifference equations on the support; small-scale problems are batch-solvable; watch for sparsity and condition number at scale.
|
|
35
|
+
- **Exact large-scale multi-agent equilibrium**: $n$-player Nash equilibrium is PPAD-complete; general case is **intractable** → **anti-pattern**; switch to approximate / learning-based methods (fictitious play, self-play, multi-agent RL, NFSP).
|
|
36
|
+
- **Evolutionary / repeated game simulation**: Population and strategy batch updates are parallelizable; watch for inter-agent communication and synchronization overhead.
|
|
37
|
+
|
|
38
|
+
Eight-dimensional minimum assessment (formal terms): **Tensorization** — whether players / strategies / rounds admit batched unfolding; **GEMM-mappability** — whether payoff matrices, minimax, LP / linear systems fall into matrix operations; **Complexity** — flag PPAD / combinatorial explosion risks for Nash / mechanism solving; **Memory and KV-Cache** — whether the strategy-state Cartesian product is compressible; **Low-precision stability** — whether equilibrium solving, LP, soft best-response have controlled condition numbers; **Parallelism and communication** — synchronization cost of multi-agent rollouts; **Sparse structure** — whether the interaction graph is structured; **Operator fusion** — whether payoff, mask, and best-response updates can be fused.
|
|
39
|
+
|
|
40
|
+
> Use in conjunction with `../../references/books/optimization-ml.md` (duality / minimax, game learning algorithms) and `../../references/books/matrix-analysis.md` (matrix games, linear systems).
|
|
41
|
+
|
|
42
|
+
## When NOT to Use
|
|
43
|
+
|
|
44
|
+
- **Single-agent decision problems with no interaction** — optimization thinking is more appropriate; no need to model others' strategic responses.
|
|
45
|
+
- **Purely cooperative problems with no conflict of interest** — players' objectives are fully aligned; strategic analysis is unnecessary.
|
|
46
|
+
- **Deterministic problems with no strategic uncertainty** — outcomes are uniquely determined by one's own actions; no reaction from others is involved.
|
|
47
|
+
- **Outcomes determined purely by chance with no strategic choice** — no strategy sets to analyze; game theory does not apply.
|
|
48
|
+
|
|
49
|
+
## When to Use
|
|
50
|
+
|
|
51
|
+
- Multiple decision-makers interact; each player's strategy affects others' payoffs; the optimal strategy depends on predicting others' responses.
|
|
52
|
+
- Need to design incentive mechanisms so that self-interested behavior leads to socially optimal outcomes; predict equilibrium states under given rules.
|
|
53
|
+
- Analyze pricing, entry, and exit in competitive markets; benefit allocation and threat points in negotiation / bargaining.
|
|
54
|
+
- **Design strategies and equilibrium objectives for multi-agent systems, adversarial training (GAN / robust learning), and routing games.**
|
|
55
|
+
|
|
56
|
+
## Method
|
|
57
|
+
|
|
58
|
+
### Step 1: Identify Players and Strategies
|
|
59
|
+
Define the fundamental elements of the game: players $N=\{1,\ldots,n\}$; strategy sets $S_i$; payoff functions $u_i: S_1\times\cdots\times S_n\to\mathbb{R}$; information structure (symmetric or not? incomplete information? sequential actions? communication possible?). The core task is identifying the "interaction structure" — whose choices affect whose outcomes. Missing a key player is the most common error.
|
|
60
|
+
|
|
61
|
+
### Step 2: Analyze the Game Type
|
|
62
|
+
First classify: zero-sum / general-sum, simultaneous / sequential, cooperative / non-cooperative, complete / incomplete information. The type determines the method: zero-sum uses minimax / LP; sequential games use backward induction and subgame-perfect equilibrium; cooperative games use the Shapley value to allocate by marginal contribution:
|
|
63
|
+
$$\phi_i(v)=\sum_{S\subseteq N\setminus\{i\}}\frac{|S|!(n-|S|-1)!}{n!}[v(S\cup\{i\})-v(S)]$$
|
|
64
|
+
Misclassifying the type leads directly to applying the wrong equilibrium concept and solution method.
|
|
65
|
+
|
|
66
|
+
### Step 3: Construct the Payoff Matrix / Function
|
|
67
|
+
For two-player games, use a payoff matrix with cells $(u_1,u_2)$; for $n$-player games, use $u_i(s_1,\ldots,s_n)$. Verify completeness: are any players or strategies missing? Do payoffs accurately reflect preferences? Are there externalities? Payoffs are the foundation of all analysis; inaccurate assessment biases everything downstream.
|
|
68
|
+
|
|
69
|
+
### Step 4: Search for Dominant Strategies
|
|
70
|
+
Strict dominance: $u_i(s_i^*,s_{-i}) > u_i(s_i,s_{-i})$ for all $s_i\neq s_i^*$ and all $s_{-i}$; weak dominance: $\geq$ with at least one strict inequality. **Iterated elimination of dominated strategies** successively removes inferior strategies to shrink the strategy space. A dominant-strategy equilibrium is the strongest possible conclusion — no player has any incentive to deviate. If none exists, proceed to equilibrium analysis.
|
|
71
|
+
|
|
72
|
+
### Step 5: Compute Equilibria
|
|
73
|
+
Select the method based on the game type from Step 2. Pure-strategy Nash: for each $s_{-i}$, find the best response $BR_i(s_{-i})$; intersections are equilibria. Mixed strategies: find probability distributions that make others indifferent, $u_j(s_j,\sigma^*_{-j})=u_j(s_j',\sigma^*_{-j})$ for all $s_j,s_j'\in S_j$. Zero-sum matrix games use minimax $\max_x\min_y f=\min_y\max_x f$; sequential games use backward induction; cooperative games output allocation schemes rather than standard Nash. For multiple equilibria, perform Pareto ranking and label optimal vs. inferior equilibria.
|
|
74
|
+
|
|
75
|
+
### Step 6: Check Equilibrium Stability
|
|
76
|
+
Trembling-hand perfection (stable under small perturbations, excluding equilibria that rely on "opponents never make mistakes"); evolutionarily stable strategy ESS (resists invasion in repeated interactions); off-equilibrium-path incentives (in sequential games, whether there is motivation to honor commitments). Ask: is the equilibrium unique or multiple? Is it robust to small perturbations? Only stable equilibria constitute credible predictions.
|
|
77
|
+
|
|
78
|
+
### Step 7: Mechanism Design and Improvement
|
|
79
|
+
If no good equilibrium exists, redesign the game: modify the payoff structure (incentives and penalties), add enforceable rules, introduce communication and reputation mechanisms, Vickrey auctions (truthful valuation reporting is a dominant strategy). Mechanism design formalization:
|
|
80
|
+
$$\max\ \text{social welfare}\quad\text{s.t.}\quad\text{incentive compatibility + individual rationality}$$
|
|
81
|
+
The revelation principle: any Bayesian game equilibrium can be implemented by a direct mechanism — truthful type reporting is an equilibrium. Mechanism design is the most practical branch — rather than lamenting non-cooperation, change the rules so that cooperation becomes the self-interested choice.
|
|
82
|
+
|
|
83
|
+
## Common Errors
|
|
84
|
+
|
|
85
|
+
| Error | Critique | Correct Approach |
|
|
86
|
+
|---|---|---|
|
|
87
|
+
| Assuming others think the same way you do | Others have different payoff functions / rationality levels; $u_i\neq u_j$ | Explicitly specify each player's payoff function and beliefs |
|
|
88
|
+
| Ignoring mixed strategies | Pure-strategy equilibria may not exist; mixed equilibria always exist (Nash, 1950) | Check whether mixed-strategy analysis is needed |
|
|
89
|
+
| Confusing Nash equilibrium with Pareto optimality | Nash equilibrium can be Pareto-dominated (Prisoner's Dilemma); the two are independent | Separately annotate equilibrium properties and efficiency properties |
|
|
90
|
+
| Ignoring information asymmetry | Incomplete-information equilibria differ fundamentally from complete-information ones | Distinguish Bayesian games from complete-information games |
|
|
91
|
+
| Oversimplifying payoff structure | Omitting key payoff dimensions distorts the equilibrium | Systematically check complete payoffs for all players |
|
|
92
|
+
| Ignoring repeated interaction dynamics | One-shot equilibria may change in repeated games (folk theorem) | Analyze subgame-perfect equilibria |
|
|
93
|
+
| Intractable exact large-scale equilibrium | $n$-player Nash equilibrium is PPAD-complete; exact solutions are infeasible | Pass through the GPU eight-dimensional gate; switch to approximate / learning methods (self-play, MARL) |
|
|
94
|
+
|
|
95
|
+
## Operating Procedure
|
|
96
|
+
|
|
97
|
+
When this skill is triggered, the output must include:
|
|
98
|
+
|
|
99
|
+
1. **[Players]**: All decision-makers and their types (rational / boundedly rational / unknown).
|
|
100
|
+
2. **[Strategy Sets]**: Available strategies for each player; label finite / infinite.
|
|
101
|
+
3. **[Payoff Matrix]**: Payoff values $(u_1,u_2,\ldots)$ for core strategy profiles, or payoff function form.
|
|
102
|
+
4. **[Game Type]**: Zero-sum / non-zero-sum / sequential / cooperative, with corresponding analytical method.
|
|
103
|
+
5. **[Equilibria]**: All applicable equilibria (pure strategy, mixed strategy, subgame-perfect, cooperative allocation, etc.); label multiple equilibria.
|
|
104
|
+
6. **[Stability]**: Trembling-hand perfection, evolutionary stability, off-equilibrium-path incentives.
|
|
105
|
+
7. **[Mechanism Recommendations]**: If equilibria are undesirable, propose mechanism design suggestions (modify payoffs, add rules, introduce reputation, etc.).
|
|
106
|
+
8. **[GPU Feasibility]**: Matrix games / minimax → LP is GEMM-compatible; large-scale multi-agent equilibria may be intractable → approximate / learning methods; pass through the eight-dimensional gate.
|
|
107
|
+
|
|
108
|
+
**Output must not consist of analysis alone without conclusions.**
|
|
109
|
+
|
|
110
|
+
## Relations to Other Skills
|
|
111
|
+
|
|
112
|
+
- **Optimization Thinking**: Nash equilibrium is mutual optimization — each player optimizes their own payoff subject to others' strategies; the equilibrium is the common solution to all optimization problems.
|
|
113
|
+
- **Probability and Statistics**: Mixed strategies select actions via probability distributions; Bayesian games handle type uncertainty probabilistically; equilibria depend on expected payoffs.
|
|
114
|
+
- **Information Theory Thinking**: Information asymmetry and signaling — in Spence signaling games, the information structure directly affects equilibrium existence and properties.
|
|
115
|
+
- **Causal Inference Thinking**: Strategic choices causally affect others' payoffs; causal chains in strategic interaction are bidirectional.
|
|
116
|
+
- **Algorithmic Thinking**: Nash equilibrium computation complexity (PPAD-complete), Lemke-Howson algorithm, evolutionary game simulation.
|
|
117
|
+
- **Modern Mathematics Activation**: `../../references/books/optimization-ml.md` (duality / minimax, game learning algorithms), `../../references/books/matrix-analysis.md` (matrix games, linear algebraic structure of payoff matrices).
|
|
@@ -1,318 +1,123 @@
|
|
|
1
|
-
---
|
|
2
|
-
name: game-theory
|
|
3
|
-
description: |
|
|
4
|
-
|
|
5
|
-
|
|
6
|
-
|
|
7
|
-
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
> "
|
|
14
|
-
>
|
|
15
|
-
>
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
|
|
30
|
-
|
|
31
|
-
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
-
|
|
51
|
-
-
|
|
52
|
-
-
|
|
53
|
-
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
-
|
|
58
|
-
-
|
|
59
|
-
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
#### 科研模式 / Research Mode
|
|
126
|
-
|
|
127
|
-
- **2人博弈**:用支付矩阵(表格)表示,行参与者和列参与者的策略组合对应单元格 $(u_1, u_2)$
|
|
128
|
-
- **n人博弈**:用效用函数 $u_i(s_1, \ldots, s_n)$ 表示,每个参与者的收益依赖于所有人的策略
|
|
129
|
-
- 验证完备性:是否遗漏了关键参与者或策略?支付是否准确反映偏好?
|
|
130
|
-
|
|
131
|
-
关键问题:支付是零和的还是一般和的?是否存在外部性(一人的策略影响非参与者)?
|
|
132
|
-
|
|
133
|
-
#### 生活模式 / Life Mode
|
|
134
|
-
|
|
135
|
-
梳理各方的得失:
|
|
136
|
-
- **各方的得失是什么?** 不同选择下每个人会得到什么、失去什么?得失不限于金钱——时间、情感、声誉都是支付
|
|
137
|
-
- **不同选择下每个人会得到什么、失去什么?** 尝试画出简单的"如果A选X、B选Y,各方的结果"
|
|
138
|
-
- **你是否完整考虑了所有人的利益?** 只考虑自己的支付而忽略他人的,会导致错误的策略预判
|
|
139
|
-
|
|
140
|
-
#### 共通要点 / Common Key Points
|
|
141
|
-
|
|
142
|
-
支付(payoff)是博弈分析的根基。如果支付评估不准,后续所有分析都会偏。务必考虑所有参与者的完整利益维度。
|
|
143
|
-
|
|
144
|
-
---
|
|
145
|
-
|
|
146
|
-
### 第三步:寻找占优策略 / Find Dominant Strategies
|
|
147
|
-
|
|
148
|
-
#### 科研模式 / Research Mode
|
|
149
|
-
|
|
150
|
-
- **严格占优策略(Strictly Dominant)**:$s_i^*$ 使得 $u_i(s_i^*, s_{-i}) > u_i(s_i, s_{-i})$ 对所有 $s_i \neq s_i^*$ 和所有 $s_{-i}$ 成立
|
|
151
|
-
- **弱占优策略(Weakly Dominant)**:$u_i(s_i^*, s_{-i}) \geq u_i(s_i, s_{-i})$ 对所有 $s_{-i}$ 成立,且至少一个 $s_{-i}$ 严格不等
|
|
152
|
-
- **迭代消去劣策略(Iterated Elimination of Dominated Strategies)**:依次删去每个参与者的劣策略,缩小策略空间
|
|
153
|
-
|
|
154
|
-
若存在占优策略均衡,这是最强形式的均衡——无人有任何偏离动机,不论理性程度。
|
|
155
|
-
|
|
156
|
-
#### 生活模式 / Life Mode
|
|
157
|
-
|
|
158
|
-
问自己:
|
|
159
|
-
- **有没有一个选择,无论对方怎么做都是你的最佳?** 如果有,直接选它——不需要猜测对方意图
|
|
160
|
-
- 如果没有占优策略,你的最优选择取决于对方怎么做——这时需要进入下一步分析
|
|
161
|
-
- 排除明显不好的选项:有些选择在任何情况下都不如其他选择,先删掉它们简化决策
|
|
162
|
-
|
|
163
|
-
#### 共通要点 / Common Key Points
|
|
164
|
-
|
|
165
|
-
占优策略是最强的结论——一旦找到,问题就解决了。如果没有占优策略,才需要更复杂的均衡分析。
|
|
166
|
-
|
|
167
|
-
---
|
|
168
|
-
|
|
169
|
-
### 第四步:计算纳什均衡 / Compute Nash Equilibria
|
|
170
|
-
|
|
171
|
-
#### 科研模式 / Research Mode
|
|
172
|
-
|
|
173
|
-
- **纯策略纳什均衡**:通过最优响应分析(best response analysis)——对每个 $s_{-i}$,找出 $i$ 的最优响应 $BR_i(s_{-i})$,交叉点即为均衡
|
|
174
|
-
- **混合策略纳什均衡**:计算使其他参与者无差异的概率分布——参与者 $i$ 的混合策略 $\sigma_i$ 使得参与者 $j$ 的所有纯策略期望支付相等
|
|
175
|
-
|
|
176
|
-
$$u_j(s_j, \sigma^*_{-j}) = u_j(s_j', \sigma^*_{-j}) \quad \text{for all } s_j, s_j' \in S_j$$
|
|
177
|
-
|
|
178
|
-
- **多重均衡**:帕累托排序——比较各均衡下所有参与者的支付,标注帕累托最优均衡与帕累托劣均衡
|
|
179
|
-
|
|
180
|
-
#### 生活模式 / Life Mode
|
|
181
|
-
|
|
182
|
-
寻找稳定状态:
|
|
183
|
-
- **是否存在一个稳定状态——大家都不愿意单方面改变策略?** 这就是最可能的结果
|
|
184
|
-
- 如果存在这样的稳定状态,那就是最可能的结局——每个人都在做自己能做的最好选择
|
|
185
|
-
- **如果有多个稳定状态,哪个更可能发生?** 多重均衡意味着结果不确定——需要看历史惯例、沟通协调或焦点效应(salience)
|
|
186
|
-
|
|
187
|
-
#### 共通要点 / Common Key Points
|
|
188
|
-
|
|
189
|
-
纳什均衡是博弈论的核心预测工具——它告诉你"在理性假设下,最可能的结果是什么"。多重均衡是常态,需要额外信息选择。
|
|
190
|
-
|
|
191
|
-
---
|
|
192
|
-
|
|
193
|
-
### 第五步:分析博弈类型 / Analyze Game Type
|
|
194
|
-
|
|
195
|
-
#### 科研模式 / Research Mode
|
|
196
|
-
|
|
197
|
-
- **零和博弈**:minimax 定理 $\max_x \min_y f(x,y) = \min_y \max_x f(x,y)$,最优策略是保守的——在最坏情况下最大化收益
|
|
198
|
-
- **序列博弈**:逆向归纳法(backward induction)与子博弈完美均衡(subgame perfect equilibrium)——每一步都是后续子博弈的纳什均衡
|
|
199
|
-
- **合作博弈**:联盟形成(coalition formation)与 Shapley 值(Shapley Value)——按边际贡献分配联盟收益
|
|
200
|
-
|
|
201
|
-
$$\phi_i(v) = \sum_{S \subseteq N \setminus \{i\}} \frac{|S|!(n - |S| - 1)!}{n!} [v(S \cup \{i\}) - v(S)]$$
|
|
202
|
-
|
|
203
|
-
#### 生活模式 / Life Mode
|
|
204
|
-
|
|
205
|
-
判断互动类型,选择对应策略:
|
|
206
|
-
- **如果是竞争关系(一方得另一方必失)**:保守策略——做最坏情况下最好的选择。不要指望对方合作,先确保自己不会太惨
|
|
207
|
-
- **如果是合作关系**:寻找双赢点——双方都能改善的选项。关注如何分配合作收益,而非如何争夺
|
|
208
|
-
- **如果是谈判**:识别双方的底线和共识区域——谈判的核心是找到双方都能接受的区域,而非各自的最优
|
|
209
|
-
|
|
210
|
-
#### 共通要点 / Common Key Points
|
|
211
|
-
|
|
212
|
-
博弈类型决定了分析方法——零和用保守策略,序列用逆向推理,合作用联盟分析。误判类型会导致方法错用。
|
|
213
|
-
|
|
214
|
-
---
|
|
215
|
-
|
|
216
|
-
### 第六步:检查均衡稳定性 / Check Equilibrium Stability
|
|
217
|
-
|
|
218
|
-
#### 科研模式 / Research Mode
|
|
219
|
-
|
|
220
|
-
- **颤抖手完美(Trembling-Hand Perfect Equilibrium)**:均衡在微小失误下仍然稳定——排除依赖"对手绝不犯错"的均衡
|
|
221
|
-
- **演化稳定性(Evolutionarily Stable Strategy, ESS)**:在重复互动中,均衡策略能否抵御入侵策略——生物学与经济学交叉
|
|
222
|
-
- **偏离均衡路径的激励(Off-Equilibrium Path Incentives)**:序列博弈中,参与者是否有动机在非均衡节点上遵守承诺
|
|
223
|
-
|
|
224
|
-
关键问题:均衡是唯一的还是多重的?多重均衡中哪个更可能被选择?均衡是否对微小扰动鲁棒?
|
|
225
|
-
|
|
226
|
-
#### 生活模式 / Life Mode
|
|
227
|
-
|
|
228
|
-
评估稳定状态的可靠性:
|
|
229
|
-
- **这个稳定状态靠谱吗?** 如果有人犯小错(误解你的意图、操作失误),局面会崩吗?
|
|
230
|
-
- **在重复互动中,这个状态能持续吗?** 单次博弈的结论在长期关系中可能完全不同——合作可以在重复互动中涌现
|
|
231
|
-
- 如果稳定状态对小错误不鲁棒,实际结局可能偏离理论预测——需要考虑容错方案
|
|
232
|
-
|
|
233
|
-
#### 共通要点 / Common Key Points
|
|
234
|
-
|
|
235
|
-
稳定的均衡才是可信的预测。不稳定的均衡可能在现实中永远不会出现。
|
|
236
|
-
|
|
237
|
-
---
|
|
238
|
-
|
|
239
|
-
### 第七步:设计机制与改进 / Design Mechanisms and Improve
|
|
240
|
-
|
|
241
|
-
#### 科研模式 / Research Mode
|
|
242
|
-
|
|
243
|
-
如果不存在好的纳什均衡,重新设计博弈:
|
|
244
|
-
- **改变支付结构**:引入惩罚或奖励,使均衡转向更优结果
|
|
245
|
-
- **增加规则**:引入可强制执行的协议、惩罚偏离者
|
|
246
|
-
- **引入通信与声誉**:重复博弈中声誉机制可维持合作
|
|
247
|
-
- **Vickrey 拍卖(第二价格拍卖)**:报真实估值是占优策略——机制设计的经典范例
|
|
248
|
-
- **机制设计(Mechanism Design)**:在满足激励相容性(incentive compatibility)和效率(efficiency)的前提下设计规则
|
|
249
|
-
|
|
250
|
-
$$\text{Mechanism Design: } \max \text{ social welfare} \quad \text{s.t.} \quad \text{incentive compatibility + individual rationality}$$
|
|
251
|
-
|
|
252
|
-
显示原理(Revelation Principle):任何贝叶斯博弈的均衡都可由一个直接机制(direct mechanism)实现——参与者如实报告类型是均衡。
|
|
253
|
-
|
|
254
|
-
#### 生活模式 / Life Mode
|
|
255
|
-
|
|
256
|
-
改变规则让好结局更自然:
|
|
257
|
-
- **如果当前规则下大家都陷入不好的结局(如囚徒困境),能否改变规则让好结局成为自然选择?** 这是机制设计的核心思想
|
|
258
|
-
- **引入承诺**:先做出不可反悔的承诺,改变他人的预期和选择(如预先公开你的底线)
|
|
259
|
-
- **声誉机制**:在长期互动中,好声誉本身就是资产——维护声誉可以维持合作
|
|
260
|
-
- **改变激励结构**:让"做好事"变得自然——调整奖惩让每个人的最优选择恰好也是整体最优
|
|
261
|
-
|
|
262
|
-
#### 共通要点 / Common Key Points
|
|
263
|
-
|
|
264
|
-
机制设计是博弈论最实用的分支——与其抱怨人不合作,不如改变规则让合作成为自利选择。
|
|
265
|
-
|
|
266
|
-
## 常见错误 / Common Errors
|
|
267
|
-
|
|
268
|
-
| 错误 / Error | 数学批评 / Mathematical Critique | 正确做法 / Correct Approach |
|
|
269
|
-
|-------------|-------------------------------|---------------------------|
|
|
270
|
-
| 假设他人像自己一样思考 `[科研/通用]` | 他人可能有不同的支付函数或理性水平,$u_i \neq u_j$ | 明确各参与者的支付函数与信念 |
|
|
271
|
-
| 忽略混合策略 `[科研]` | 纯策略均衡可能不存在,但混合策略均衡总存在(Nash, 1950) | 检查是否需要混合策略分析 |
|
|
272
|
-
| 混淆纳什均衡与帕累托最优 `[科研/通用]` | 纳什均衡可以是帕累托劣的(囚徒困境),两者独立 | 分别标注均衡性与效率性 |
|
|
273
|
-
| 忽略信息不对称 `[科研]` | 不完全信息博弈的均衡与完全信息博弈截然不同 | 区分贝叶斯博弈与完全信息博弈 |
|
|
274
|
-
| 过度简化支付结构 `[科研/通用]` | 遗漏关键支付维度使均衡分析失真 | 系统检查所有参与者的完整支付 |
|
|
275
|
-
| 忽略重复互动动态 `[科研/通用]` | 单次博弈的均衡在重复博弈中可能改变(folk theorem) | 分析重复博弈的子博弈完美均衡 |
|
|
276
|
-
| 假设别人和你想法一样 `[生活]` | 不同人有不同动机、偏好和信息,不能用自己的思维代入他人 | 主动了解他人的真实利益和约束 |
|
|
277
|
-
| 忽略情绪和非理性因素 `[生活]` | 博弈论假设理性,但现实中愤怒、信任、面子等深刻影响选择 | 识别关键情绪因素,修正支付评估 |
|
|
278
|
-
| 只考虑自己的最优不考虑他人的反应 `[生活]` | 单方优化不是博弈分析——你的最优取决于他人的选择 | 先想"如果我这么做,对方会怎么反应" |
|
|
279
|
-
|
|
280
|
-
## 操作规程 / Operating Procedure
|
|
281
|
-
|
|
282
|
-
当本 skill 被触发时,先判断模式:
|
|
283
|
-
|
|
284
|
-
> **模式选择 / Mode Selection**
|
|
285
|
-
> - 若问题涉及数学建模、论文分析、算法设计、均衡计算 → **科研模式**
|
|
286
|
-
> - 若问题涉及日常决策、人际互动、生活规划、冲突协调 → **生活模式**
|
|
287
|
-
> - 若问题同时涉及两者 → **双模式并行,生活模式优先输出,科研模式作为补充验证**
|
|
288
|
-
|
|
289
|
-
### 科研模式输出格式 / Research Mode Output Format
|
|
290
|
-
|
|
291
|
-
1. **[参与者]:[列表]**:明确所有决策者及其类型(理性/有限理性/未知)
|
|
292
|
-
2. **[策略集]:[描述]**:列出每个参与者的可选策略,标注是否为有限/无限策略集
|
|
293
|
-
3. **[支付矩阵]:[关键值]**:写出核心策略组合的支付值 $(u_1, u_2, \ldots)$ 或支付函数形式
|
|
294
|
-
4. **[均衡]:[纳什均衡策略/混合概率]**:计算所有纳什均衡(纯策略和混合策略),标注多重均衡
|
|
295
|
-
5. **[博弈类型]:[零和/非零和/序列/合作]**:分类博弈结构,选择对应分析方法
|
|
296
|
-
6. **[稳定性]:[分析]**:检查均衡的颤抖手完美性、演化稳定性、偏离均衡路径激励
|
|
297
|
-
7. **[机制建议]:[改进]**:若均衡不理想,提出机制设计建议——改变支付、增加规则、引入声誉等
|
|
298
|
-
|
|
299
|
-
**输出必须包含以上 7 项,不得只输出分析性文字而不给出结论。**
|
|
300
|
-
|
|
301
|
-
### 生活模式输出格式 / Life Mode Output Format
|
|
302
|
-
|
|
303
|
-
1. **[参与者]:[谁在影响你的决策]** — 列出所有相关方及其可能的行动
|
|
304
|
-
2. **[得失]:[各方的利益]** — 不同选择下每个人的得失
|
|
305
|
-
3. **[稳定状态]:[分析]** — 是否存在没人愿意单方面改变的策略组合?标注多重稳定状态
|
|
306
|
-
4. **[互动类型]:[竞争/合作/谈判]** — 分类互动结构,选择对应策略
|
|
307
|
-
5. **[稳定性]:[评估]** — 这个稳定状态是否对小错误鲁棒?能否在重复互动中持续?
|
|
308
|
-
6. **[规则建议]:[改进]** — 如果当前规则下结局不理想,如何改变规则让好结局更自然
|
|
309
|
-
|
|
310
|
-
**输出必须包含以上 6 项,不得只输出分析性文字而不给出结论。**
|
|
311
|
-
|
|
312
|
-
## 与其他 skill 的关系
|
|
313
|
-
|
|
314
|
-
- **优化思想**:Nash均衡是互相优化——每个参与者在他人的策略约束下优化自己的支付,均衡是所有优化问题的共同解
|
|
315
|
-
- **概率与统计**:混合策略以概率分布选择行动,贝叶斯博弈用概率处理类型不确定性,均衡计算依赖期望支付
|
|
316
|
-
- **信息论思想**:信息不对称与信号传递——Spence 信号博弈模型中,信息结构直接影响均衡的存在与性质
|
|
317
|
-
- **因果推断思想**:因果与博弈互动——参与者的策略选择因果影响他人收益,战略互动中的因果链是双向的
|
|
318
|
-
- **算法思想**:博弈算法与均衡计算——纳什均衡的计算复杂度(PPAD-complete),Lemke-Howson 算法,演化博弈的模拟算法
|
|
1
|
+
---
|
|
2
|
+
name: game-theory
|
|
3
|
+
description: |
|
|
4
|
+
触发:当策略取决于他人选择、需计算纳什均衡、分析零和/非零和博弈、做机制设计;或为多智能体系统、对抗训练、路由博弈设计策略时调用。
|
|
5
|
+
English: Trigger when optimal strategy depends on others' choices, needing Nash equilibrium computation, zero-sum/non-zero-sum analysis, mechanism design; or designing strategies for multi-agent systems, adversarial training, routing games.
|
|
6
|
+
---
|
|
7
|
+
|
|
8
|
+
> **语言路由**:若用户消息为英文,请读取并遵循同目录下的 `SKILL.en.md`,按其操作规程以英文输出;中文消息则继续使用本文件。
|
|
9
|
+
|
|
10
|
+
# 🎯 博弈论思想 / Game Theory
|
|
11
|
+
|
|
12
|
+
> "你的最优选择取决于他人的选择——思考不仅要深入,还要互动。"
|
|
13
|
+
> "Your optimal choice depends on others' choices — thinking must be interactive, not just deep."
|
|
14
|
+
>
|
|
15
|
+
> —— 博弈论、决策论、机制设计 / Game Theory, Decision Theory, Mechanism Design
|
|
16
|
+
|
|
17
|
+
## 核心原则 / Core Principle
|
|
18
|
+
|
|
19
|
+
**战略互动——你的最优行动取决于他人做什么,而他人的最优行动取决于你做什么。纳什均衡:无人能从单方面偏离中获益的策略组合。**
|
|
20
|
+
|
|
21
|
+
**Strategic interaction — your best action depends on what others do, and theirs depend on what you do. Nash equilibrium: strategy profile where no player benefits from unilateral deviation.**
|
|
22
|
+
|
|
23
|
+
关键概念:**支付矩阵(Payoff Matrix)**、**占优策略(Dominant Strategy)**、**混合策略(Mixed Strategy)**、**帕累托最优性(Pareto Optimality)**。分类维度:合作 vs 非合作、零和 vs 一般和、同时 vs 序列。
|
|
24
|
+
|
|
25
|
+
> **数学形式化 / Mathematical Formalization**
|
|
26
|
+
>
|
|
27
|
+
> 形式化定义:n 个参与者,每人有策略集 $S_i$ 和支付函数 $u_i: S_1 \times \cdots \times S_n \to \mathbb{R}$。
|
|
28
|
+
>
|
|
29
|
+
> 纳什均衡:策略组合 $s^* = (s_1^*, \ldots, s_n^*)$ 满足 $u_i(s^*) \geq u_i(s_i, s^*_{-i})$ 对所有 $s_i \in S_i$、所有 $i$ 成立,其中 $s^*_{-i}$ 为除 $i$ 外所有人的策略。混合策略纳什均衡:概率分布 $\sigma^*$ 使 $u_i(\sigma^*) \geq u_i(s_i, \sigma^*_{-i})$。
|
|
30
|
+
>
|
|
31
|
+
> 占优策略:严格 $u_i(s_i^*, s_{-i}) > u_i(s_i, s_{-i})$;弱占优 $\geq$ 且至少一处严格。
|
|
32
|
+
>
|
|
33
|
+
> 零和 minimax:$\max_x \min_y f(x,y) = \min_y \max_x f(x,y)$。
|
|
34
|
+
|
|
35
|
+
## GPU 友好性 / GPU-Friendliness(横切检查)
|
|
36
|
+
|
|
37
|
+
博弈求解映射 GPU 的能力差异极大,过 `../../references/gpu-friendly-math.md` 八维门:
|
|
38
|
+
|
|
39
|
+
- **矩阵博弈 / 双人零和(minimax)**:等价于线性规划(LP),支付矩阵运算可 GEMM 化、可批量并行——**友好**。
|
|
40
|
+
- **混合策略均衡求解**:支撑集上的线性方程组/无差异方程,小规模可批量解;规模大时注意稀疏性与条件数。
|
|
41
|
+
- **大规模多智能体均衡精确求解**:n 人纳什均衡 PPAD-complete,一般情形**不可算**→ **反模式**,改用近似/学习方法(fictitious play、虚拟自博、多智能体 RL、NFSP)。
|
|
42
|
+
- **演化/重复博弈模拟**:群体与策略批量更新可并行;注意 agent 间通信与同步开销。
|
|
43
|
+
|
|
44
|
+
八维最低判定(正式术语):**张量化**看玩家/策略/回合是否能批量展开;**GEMM 可映射**看支付矩阵、minimax、LP/线性方程是否落矩阵运算;**复杂度**标注 Nash/机制求解的 PPAD/组合爆炸风险;**显存与 KV-Cache**看策略-状态笛卡尔积是否可压缩;**低精度稳定**看均衡求解、LP、soft best-response 是否条件数可控;**并行与通信**看多智能体 rollout 的同步成本;**稀疏结构**看交互图是否结构化;**算子融合**看收益、mask、best-response 更新能否融合。
|
|
45
|
+
|
|
46
|
+
> 配合 `../../references/books/optimization-ml.md`(对偶/minimax、博弈学习算法)与 `../../references/books/matrix-analysis.md`(矩阵博弈、线性方程组)。
|
|
47
|
+
|
|
48
|
+
## 不适用场景 / When NOT to Use
|
|
49
|
+
|
|
50
|
+
- **单人决策问题,与他人无互动**——优化思想更适用,无需考虑他人策略响应。
|
|
51
|
+
- **纯合作问题,无利益冲突**——参与者目标完全一致,无需战略分析。
|
|
52
|
+
- **确定性问题,无战略不确定性**——结果由自身行动唯一决定,不涉及他人反应。
|
|
53
|
+
- **纯粹运气决定的结果,各方无策略选择**——无策略集可分析,博弈论不适用。
|
|
54
|
+
|
|
55
|
+
## 何时使用 / When to Use
|
|
56
|
+
|
|
57
|
+
- 多个决策者相互影响,各自策略影响他人收益;最佳策略取决于他人响应,需预判他人行动。
|
|
58
|
+
- 需设计激励机制,使自利行为导向社会最优结果;预测给定规则下的均衡稳定状态。
|
|
59
|
+
- 分析竞争市场中的定价、进入、退出;谈判/讨价还价中的利益分配与威胁点。
|
|
60
|
+
- **为多智能体系统、对抗训练(GAN/鲁棒学习)、路由博弈设计策略与均衡目标**。
|
|
61
|
+
|
|
62
|
+
## 方法流程 / Method
|
|
63
|
+
|
|
64
|
+
### 第一步:识别参与者与策略
|
|
65
|
+
定义博弈基本要素:参与者 $N=\{1,\ldots,n\}$;策略集 $S_i$;支付函数 $u_i: S_1\times\cdots\times S_n\to\mathbb{R}$;信息结构(是否对称?是否不完全信息?行动是否序列?能否通信?)。核心是识别"互动结构"——谁的选择影响谁的结果,遗漏关键参与者是最常见错误。
|
|
66
|
+
|
|
67
|
+
### 第二步:分析博弈类型
|
|
68
|
+
先判定零和/一般和、同时/序列、合作/非合作、完全信息/不完全信息。类型决定方法:零和用 minimax/LP;序列博弈用逆向归纳与子博弈完美均衡;合作博弈用 Shapley 值按边际贡献分配:
|
|
69
|
+
$$\phi_i(v)=\sum_{S\subseteq N\setminus\{i\}}\frac{|S|!(n-|S|-1)!}{n!}[v(S\cup\{i\})-v(S)]$$
|
|
70
|
+
误判类型会直接导致均衡概念和求解方法错用。
|
|
71
|
+
|
|
72
|
+
### 第三步:构建支付矩阵/函数
|
|
73
|
+
2 人博弈用支付矩阵,单元格 $(u_1,u_2)$;n 人博弈用 $u_i(s_1,\ldots,s_n)$。验证完备性:是否遗漏参与者或策略?支付是否准确反映偏好?是否存在外部性?支付是分析根基,评估不准则后续全偏。
|
|
74
|
+
|
|
75
|
+
### 第四步:寻找占优策略
|
|
76
|
+
严格占优 $u_i(s_i^*,s_{-i}) > u_i(s_i,s_{-i})$ 对所有 $s_i\neq s_i^*$、所有 $s_{-i}$;弱占优 $\geq$ 且至少一处严格;**迭代消去劣策略**依次删去劣策略缩小空间。存在占优策略均衡是最强结论——无人有任何偏离动机;否则进入均衡分析。
|
|
77
|
+
|
|
78
|
+
### 第五步:计算均衡
|
|
79
|
+
按第二步的博弈类型选方法。纯策略纳什:对每个 $s_{-i}$ 找最优响应 $BR_i(s_{-i})$,交叉点即均衡。混合策略:求使他人无差异的概率分布 $u_j(s_j,\sigma^*_{-j})=u_j(s_j',\sigma^*_{-j})$ 对所有 $s_j,s_j'\in S_j$。零和矩阵博弈可走 minimax $\max_x\min_y f=\min_y\max_x f$;序列博弈用逆向归纳;合作博弈输出分配方案而非普通纳什。多重均衡做帕累托排序,标注最优与劣均衡。
|
|
80
|
+
|
|
81
|
+
### 第六步:检查均衡稳定性
|
|
82
|
+
颤抖手完美(微小失误下仍稳定,排除依赖"对手绝不犯错"的均衡);演化稳定策略 ESS(重复互动中抵御入侵);偏离均衡路径激励(序列博弈中是否有动机遵守承诺)。追问:均衡唯一还是多重?对微小扰动是否鲁棒?稳定均衡才是可信预测。
|
|
83
|
+
|
|
84
|
+
### 第七步:设计机制与改进
|
|
85
|
+
若不存在好的均衡,重新设计博弈:改变支付结构(奖惩)、增加可执行规则、引入通信与声誉、Vickrey 拍卖(报真实估值为占优策略)。机制设计形式化:
|
|
86
|
+
$$\max\ \text{social welfare}\quad\text{s.t.}\quad\text{incentive compatibility + individual rationality}$$
|
|
87
|
+
显示原理:任何贝叶斯博弈均衡都可由直接机制实现——如实报告类型即均衡。机制设计是最实用分支——与其抱怨人不合作,不如改变规则让合作成为自利选择。
|
|
88
|
+
|
|
89
|
+
## 常见错误 / Common Errors
|
|
90
|
+
|
|
91
|
+
| 错误 / Error | 批评 / Critique | 正确做法 / Correct Approach |
|
|
92
|
+
|---|---|---|
|
|
93
|
+
| 假设他人像自己一样思考 | 他人支付函数/理性水平不同,$u_i\neq u_j$ | 明确各参与者支付函数与信念 |
|
|
94
|
+
| 忽略混合策略 | 纯策略均衡可能不存在,混合均衡总存在(Nash, 1950) | 检查是否需混合策略分析 |
|
|
95
|
+
| 混淆纳什均衡与帕累托最优 | 纳什均衡可帕累托劣(囚徒困境),两者独立 | 分别标注均衡性与效率性 |
|
|
96
|
+
| 忽略信息不对称 | 不完全信息均衡与完全信息截然不同 | 区分贝叶斯博弈与完全信息博弈 |
|
|
97
|
+
| 过度简化支付结构 | 遗漏关键支付维度使均衡失真 | 系统检查所有参与者完整支付 |
|
|
98
|
+
| 忽略重复互动动态 | 单次均衡在重复博弈中可能改变(folk theorem) | 分析子博弈完美均衡 |
|
|
99
|
+
| 大规模均衡精确求解不可算 | n 人纳什均衡 PPAD-complete,精确解不可行 | 过 GPU 八维门,改用近似/学习(虚拟自博、MARL) |
|
|
100
|
+
|
|
101
|
+
## 操作规程 / Operating Procedure
|
|
102
|
+
|
|
103
|
+
当本 skill 被触发时,输出必须包含:
|
|
104
|
+
|
|
105
|
+
1. **[参与者]**:所有决策者及其类型(理性/有限理性/未知)。
|
|
106
|
+
2. **[策略集]**:每个参与者的可选策略,标注有限/无限。
|
|
107
|
+
3. **[支付矩阵]**:核心策略组合的支付值 $(u_1,u_2,\ldots)$ 或支付函数形式。
|
|
108
|
+
4. **[博弈类型]**:零和/非零和/序列/合作,对应分析方法。
|
|
109
|
+
5. **[均衡]**:所有适用均衡(纯策略、混合策略、子博弈完美、合作分配等),标注多重均衡。
|
|
110
|
+
6. **[稳定性]**:颤抖手完美性、演化稳定性、偏离均衡路径激励。
|
|
111
|
+
7. **[机制建议]**:若均衡不理想,提出机制设计建议(改变支付、增加规则、引入声誉等)。
|
|
112
|
+
8. **[GPU 可行性]**:矩阵博弈/minimax→LP 可 GEMM;大规模多智能体均衡是否不可算→近似/学习方法,过八维门。
|
|
113
|
+
|
|
114
|
+
**输出不得只给分析而无结论。**
|
|
115
|
+
|
|
116
|
+
## 与其他 skill 的关系 / Relations to Other Skills
|
|
117
|
+
|
|
118
|
+
- **优化思想**:纳什均衡是互相优化——每人在他人策略约束下优化自己支付,均衡是所有优化问题的共同解。
|
|
119
|
+
- **概率与统计**:混合策略以概率分布选行动,贝叶斯博弈用概率处理类型不确定性,均衡依赖期望支付。
|
|
120
|
+
- **信息论思想**:信息不对称与信号传递——Spence 信号博弈中信息结构直接影响均衡存在与性质。
|
|
121
|
+
- **因果推断思想**:策略选择因果影响他人收益,战略互动中的因果链是双向的。
|
|
122
|
+
- **算法思想**:纳什均衡计算复杂度(PPAD-complete)、Lemke-Howson 算法、演化博弈模拟。
|
|
123
|
+
- **现代数学激活**:`../../references/books/optimization-ml.md`(对偶/minimax、博弈学习算法)、`../../references/books/matrix-analysis.md`(矩阵博弈、支付矩阵的线性代数结构)。
|