Technical Summary: Parametric Open-Source Games

Problem Statement

Open-source game theory traditionally investigates settings where agents can inspect one another's internal decision procedures. In classical frameworks (e.g., Tennenholtz, 2004; Fortnow, 2009), agents submit symbolic programs that may condition their actions on the opponent's code. While this paradigm allows for outcomes distinct from standard Nash equilibria—such as conditional cooperation or punishment—it relies heavily on discrete, symbolic, or proof-theoretic models.

Conversely, modern multi-agent learning systems are typically continuous, differentiable, and parameterized by high-dimensional vectors (e.g., deep neural networks). Existing differentiable game literature generally assumes that a player's policy depends only on their own parameters, ignoring the possibility of conditioning on an opponent's internal state. There is a gap in the literature for a model that preserves the core "open-source" feature (behavior dependent on an opponent's internal description) while remaining compatible with standard continuous optimization tools used in modern machine learning.

Methodology

The authors introduce Parametric Open-Source Games, a continuous analogue to program equilibria.

1. Formal Framework:

   • The setting is a finite normal-form game $G = (N, (S_i), (u_i))$.
   • Each player $i$ selects a parameter vector $\theta_i$ from a compact convex set $\Theta_i$.
   • A semantics map $\phi_i: \Theta \to \Delta(S_i)$ is defined, which is a continuous function converting the full parameter profile $\theta = (\theta_1, \dots, \theta_n)$ into a mixed action for player $i$.
   • The induced payoff is $U_i(\theta) = u_i(\phi_1(\theta), \dots, \phi_n(\theta))$.
   • The game is open-source if $\phi_i$ depends on $\theta_{-i}$ (opponents' parameters); otherwise, it is closed-source.
   • A Parametric Program Nash Equilibrium (PPNE) is defined as a profile $\theta^*$ where no player can unilaterally improve their induced payoff by changing their parameter vector.

2. Semantics Families:

   • Sigmoid Semantics: In symmetric $2 \times 2$ games with actions $\{C, D\}$, the probability of cooperation is modeled as $p_i(\theta) = \sigma(\theta_i + \gamma \theta_{-i})$, where $\sigma$ is the sigmoid function and $\gamma \geq 0$ represents the coupling strength (open-source dependence).
   • Neural Semantics: To allow for richer, non-linear dependencies while preserving interpretability, the authors propose $p_i(\theta) = \sigma(\alpha \theta_i + \beta \theta_{-i} + h(\theta; W))$. Here, $\alpha$ and $\beta$ are learnable scalars representing sensitivity to self and opponent parameters, respectively. The residual network $h$ adds non-linearity but is constrained such that its first-order derivative vanishes at the origin, ensuring the local incentive structure is governed by the ratio $\beta/\alpha$.

3. Analysis Tools:

   • Learning Dynamics: The authors analyze selfish projected gradient ascent ($\theta_i^{t+1} = \Pi[\theta_i^t + \eta \nabla_{\theta_i} U_i]$).
   • Equilibrium Verification: Since gradient ascent in these models often converges to the boundary of the parameter space, the authors develop a one-dimensional boundary test (Theorem 4) to verify if a boundary point (e.g., $(B, B)$) is a PPNE by checking if unilateral deviations are profitable.

Key Contributions and Results

1. Existence and Formalization

The paper establishes that the induced parametric game is continuous over a compact domain, guaranteeing the existence of a mixed-strategy Nash equilibrium via Glicksberg's theorem. Furthermore, under quasiconcavity conditions, a pure PPNE exists via Kakutani's theorem.

2. Phase Transition in Sigmoid Semantics

For symmetric $2 \times 2$ games (Prisoner's Dilemma, Stag Hunt, Chicken, Harsh PD), the authors derive an exact coupling threshold $\gamma^*$:
$$\gamma^* = \frac{T + P - R - S}{R + T - P - S}$$
where $R, S, T, P$ are the standard payoffs for mutual cooperation, exploitation, temptation, and mutual defection.

• Result: If the coupling $\gamma < \gamma^*$, selfish gradient ascent near the origin ($\theta=0$) moves toward lower cooperation (defection). If $\gamma > \gamma^*$, the dynamics shift toward higher cooperation.
• Empirical Verification: Experiments confirm that the empirical transition in mean cooperation probability aligns closely with the analytical $\gamma^*$ across four canonical games.

3. Boundary Equilibrium Characterization

The paper provides a necessary and sufficient condition for boundary points to be PPNEs. Specifically, for a candidate boundary $(B, B)$, it is a PPNE if the derivative of the payoff with respect to a player's parameter is non-negative everywhere on the interval $(-B, B)$ when the opponent is fixed at $B$.

• Finding: In the Prisoner's Dilemma, open-source semantics with sufficient coupling shifts the best response to the cooperative boundary ($B, B$), whereas closed-source semantics shifts it to the defective boundary ($-B, -B$).

4. Neural Semantics and Sensitivity Ratios

The study extends the findings to neural semantics.

• Local Criterion: The local incentive for cooperation is governed by the ratio of cross-player sensitivity to self-player sensitivity ($\beta/\alpha$). Cooperation becomes locally attractive when $\beta/\alpha > \gamma^*$.
• Optimization Barrier: Experiments with jointly learned neural semantics reveal that while the model class can represent cooperative open-source dependence, gradient-based learning often fails to discover it from a "cold start" (initialized at $\alpha=\beta=0$). However, "warm starts" initialized above the threshold successfully converge to high-welfare cooperative regimes. This suggests the threshold acts as an optimization barrier, not just a representational limit.

Significance and Claims

The paper claims that parametric open-source games provide a bridge between symbolic program equilibria and continuous differentiable optimization. The primary significance lies in demonstrating that:

1. Qualitative Reshaping: Access to internal parameterizations (open-source coupling) can qualitatively reshape learning dynamics and equilibrium structures, steering selfish optimization toward cooperative outcomes in games where standard Nash equilibria predict defection.
2. Analytical Tractability: Open-source strategic reasoning need not be tied to symbolic representations; once encoded via a continuous semantics map, these effects can be analyzed using standard tools from continuous optimization.
3. Sensitivity Ratio: The mechanism driving cooperation is the relative sensitivity of an agent's policy to an opponent's parameters versus its own. This holds true for both simple linear couplings and complex neural semantics.

The authors maintain a modest scope, noting that the models are idealized (full parameter access, symmetric two-player, one-shot games). They posit that this framework serves as an interpretable starting point for understanding how transparency over internal representations affects incentives in parametric multi-agent systems, rather than claiming immediate applicability to complex, real-world sequential environments.