Scholarlink

<p>Interpretable strategy generation has gained significant attention for robust decision-making in adversarial environments. Existing techniques, such as indirect strategy explanation through imitation learning or strategy optimization, often suffer from performance loss during model transformation, while direct deduction methods based on logical frameworks struggle to search effectively in high-dimensional state spaces. To address these limitations, this paper introduces a neural-guided program search framework for generating interpretable strategies. Our approach integrates a Domain-Specific Language (DSL) for strategy sketching with a reinforcement learning process based on Proximal Policy Optimization (PPO), which iteratively samples from standard policies to guide program synthesis. The output of the algorithm is a set of interpretable program instructions that collectively form a task-specific strategy. Experiments in the Google Football environment demonstrate that our method outperforms several baseline approaches in terms of winning rate, number of shots, and shooting accuracy. Ablation studies further confirm the contribution of both input augmentation and strategy sketches to the robustness and interpretability of the generated strategies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable adversarial strategy synthesis with neural-guided program search

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Qian Che,
  • Yanping Xu,
  • Yijing Wang,
  • Qun Wang,
  • Wanyuan Wang,
  • Weiwei Wu

摘要

Interpretable strategy generation has gained significant attention for robust decision-making in adversarial environments. Existing techniques, such as indirect strategy explanation through imitation learning or strategy optimization, often suffer from performance loss during model transformation, while direct deduction methods based on logical frameworks struggle to search effectively in high-dimensional state spaces. To address these limitations, this paper introduces a neural-guided program search framework for generating interpretable strategies. Our approach integrates a Domain-Specific Language (DSL) for strategy sketching with a reinforcement learning process based on Proximal Policy Optimization (PPO), which iteratively samples from standard policies to guide program synthesis. The output of the algorithm is a set of interpretable program instructions that collectively form a task-specific strategy. Experiments in the Google Football environment demonstrate that our method outperforms several baseline approaches in terms of winning rate, number of shots, and shooting accuracy. Ablation studies further confirm the contribution of both input augmentation and strategy sketches to the robustness and interpretability of the generated strategies.