Game-Theoretic Objective Planning

Two approaches:

1. Actions: Turn left, go straight, turn right
2. Encode MPC into the actions

Takeaways

1. Balance between Aggressiveness (performance) and Safety
   1. So you have this balance between these two objectives.

Questions 14-dimensional cost function for the neural network

They use kind of the idea of Deep CFR, since calculating $R^T(I,a)$ takes a long time, so they train a MLP I assume to learn that?

Four contributions

1. We propose a dimension reduction formulation that better discretizes an agent’s action space
2. We display an efficient pipeline that optimizes agents’ parameters and constructs a Pareto front consisting of optimized agents across multiple objectives
3. We propose an algorithm that predicts a racing agent’s counterfactual regret against different optimized agents
4. Lastly, we provide a novel game-theoretic planning strategy that selects racing parameters using the prediction model

They use Abstraction through scoring rollout.

$o_{agg}=S_e-S_0$

$S$ is the curvilinear distance along the global race line. $S_e$ is for the ego vehicle $S_O$ is the opponent

Evolution-Based Optimization.

Concepts

• Pareto Front
• CMA-ES
• Determinantal Point Process (DPP)