Hiroki Naganuma

Bi-Level Optimization

長沼大樹 / Hiroki Naganuma

2026年10月の再サーベイ： 研究史・重要文献・最近の進展を整理した新版を追加しました（2026年10月3日基準）。以下は従来のメモです。定義・適用条件の訂正は新版を参照してください。

Algorithms for Solving Bilevel Optimization Problems

Implicit Function Theorem (in Stackelberg Games)

Introduction

• The theorem plays a pivotal role in defining solutions for the Stackelberg game.
• Provides intuition for finding the strict local Stackelberg equilibrium using gradient descent.

Key Insights

• The follower y responds optimally to the leader’s decision, making y in the upper-level problem a function of x.
• The objective function in the upper level can be expressed as g(x), independent of y.
• Using the theorem, the derivative of g w.r.t. x can be expressed in terms of derivatives of f w.r.t. x and y.
• The theorem’s application allows the leader to update its state using the derivative of g derived from the theorem, while the follower updates with gradient descent.

Research Findings

• In zero-sum games with deterministic updates, the dynamics converge only to Stackelberg equilibria with a local convergence rate.
• For general-sum games, convergence to local Stackelberg equilibrium isn’t guaranteed.
• The theorem’s direct implementation involves the leader updating its state using the derivative of g derived from the theorem.

References

• “Implicit Learning Dynamics in Stackelberg Games: Equilibria Characterization, Convergence Analysis, and Empirical Study”
• “On Solving Minimax Optimization Locally: A Follow-the-Ridge Approach”
• “Finding and Only Finding Differential Nash Equilibria by Both Pretending to be a Follower”

Penalty Based Method

• BOME! Bilevel Optimization Made Easy: A Simple First-Order Approach

Overview of the Penalty-Based Method

1. Definition:
   • A method that transforms a bilevel optimization problem into a single-objective minimization problem.
2. Details of the Method:
   • Initial Approach:
     • Explained based on the Stackelberg game.
   • Transformation:
     • First, it’s converted into a constrained optimization problem. Then, using the Lagrange multiplier, it’s transformed into an unconstrained problem.
   • Optimality Condition:
     • The function h describes the violation of the optimality condition. f2* represents the optimal f2 value.
   • Updates:
     • The values of x and y are updated by a significant delta (Δ). The choice of Δ should not deviate significantly from the gradient of f1.
3. Advantages:
   • Stabilizes the learning dynamics in games.
   • Allows the application of adaptive optimization methods.
4. Disadvantages:
   • Inefficient: Requires many gradient descent steps for each update.
   • In cases with poor loss landscapes, the optimal f2* found in each iteration can vary significantly.

Notations:

• f1: The primary objective function in the bilevel optimization problem.
• f2: The secondary objective function, which is nested within the primary f1 function and is dependent on the decisions made in f1.

Two Time Scale Update Rule (TTUR)

• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium
  • also propose FID score

Application

Hyperparameter Optimization

• Stability and Generalization of Bilevel Programming in Hyperparameter Optimization
• DARTS: Differentiable Architecture Search

Model-based Reinforcement Learning

• Model-based Reinforcement Learning: A Survey

GANs

Latent Diffusion Models

• High-Resolution Image Synthesis with Latent Diffusion Models

IRM

• Invariant Risk Minimization / Slide
• The Missing Invariance Principle Found – the Reciprocal Twin of Invariant Risk Minimization
• What Is Missing in IRM Training and Evaluation? Challenges and Solutions

Misc

• A Gradient Method for Multilevel Optimization
• Investigating Bi-Level Optimization for Learning and Vision from a Unified Perspective: A Survey and Beyond