Abstract:This paper addresses the existence of equilibria for ResNet type Mean Field Games problems. We introduce a relaxed Lagrangian approach on the Wasserstein space to handle the non-convex optimization problem of mean field type. We prove the existence of new relaxed Nash equilibria and we show that our existence result encompasses the classical Mean Field Games problem's existence result under convex data conditions.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is about the existence of equilibrium points in ResNet - type Mean Field Games (MFG) problems. Specifically, the author introduced a relaxed Lagrangian method on the Wasserstein space to deal with non - convex optimization problems and proved the existence of new relaxed Nash equilibrium points. Moreover, they showed that under the classical convex data conditions, their existence results cover the existence results of the classical Mean Field Games problems.
### Core problems and solutions of the paper
1. **Core problems**:
- In the context of deep learning (especially Residual Networks ResNet), how to deal with non - convex optimization problems in dynamic systems involving a large number of agents.
- How to ensure the existence of equilibrium solutions (i.e., Nash equilibrium) in these complex systems, even in non - convex cases.
2. **Solutions**:
- **Introducing the relaxed Lagrangian method**: By transforming the original non - convex optimization problem into an optimization problem on the probability measure space, the convexity assumption on the control variables is relaxed.
- **Proving the existence of new equilibrium points**: Using this relaxation method, the author proved that equilibrium points can still be found under broader conditions (including non - convex cases).
- **Covering the classical results**: Further proved that under the classical convex data conditions, the new relaxed equilibrium points are consistent with the traditional MFG equilibrium points.
### Explanation of mathematical formulas
To better understand the method of the paper, let's look at some key mathematical formulas:
- **State equation**:
\[
\dot{\gamma}(t)=\int_{\mathbb{R}^n} f(\gamma(t), u, m(t)) \mu_t(du)
\]
where $\gamma(t)$ is the state of each agent, $f$ is the state - transition function, $m(t)$ is the distribution of all agents, and $\mu_t$ is the strategy measure.
- **Cost function**:
\[
J_m(\gamma, \mu)=\int_0^T \int_{\mathbb{R}^n} L(\gamma(t), u, m(t)) \mu(dt, du)
\]
Here, $L$ is the Lagrangian function, representing the cost at each step.
- **Optimization problem**:
\[
\inf_P J((e_t^\sharp \pi_1^\sharp P)_{t\in[0,T]}, P)
\]
where $P$ is a probability measure such that any state - measure pair $(\gamma, \mu)\in\text{spt}(P)$ satisfies the state equation.
### Summary
By introducing the relaxed Lagrangian method, the author successfully solved the non - convex optimization challenges in ResNet - type Mean Field Games problems and proved the existence of new relaxed Nash equilibrium points. This not only expands the application scope of the traditional MFG theory but also provides new tools for dealing with optimization problems in complex dynamic systems.