毕业生知识库

足球作为世界上最热门的运动之一，向球迷传递着运动员的合作精神与顽强意志。同时，足球比赛具有的时空跨度大、行为连续高动态、要素多元耦合等特性，也给各领域研究者提出了众多科研挑战。态势信息表达了当前比赛状态以及未来发展趋势，能够有效地辅助决策执行，因此受到了广泛关注。传统态势分析方法多为面向赛后复盘的静态统计分析，缺乏面向对抗推演的动态分析，因此难以辅助实时决策过程。近年来，足球赛事数据量的增长及机器学习等理论方法的发展为足球比赛态势分析提供了新的解决方案。然而，足球比赛本身的进球稀疏、高度不确定等特性及高质量大规模数据难以获取问题都给基于数据的足球比赛态势分析方法带来了巨大挑战。

本文以足球比赛为背景，旨在研究态势智能分析方法，以及基于态势分析辅助的决策学习方法。本文的主要工作内容与创新点如下：

（1）针对足球比赛动态、不确定等特性挑战，本文基于大规模数据和深度学习方法，提出了一种基于进球期望的态势分析方法（All for Goals, AFG）。首先，对原始足球数据进行自动化时序标注，然后基于深度学习方法训练进球期望预测模型，更进一步，将进球期望概念延伸至状态评估、有球动作评估及无球跑动评估，最终完成对任意比赛状态、球员动作关于进球概率的量化。实验表明，AFG对于进球的预测在仿真环境与现实比赛中均达到了高召回率；同时，AFG在实时进球预测、事件信誉分配及无球跑动分析等任务中得到的分析结果与领域知识展示出一致性。

（2）针对态势分析结果受比赛风格、球队实力差异影响以及现实大规模高质量足球赛事数据难以获取的问题，本文提出了一种基于预训练-精调的足球比赛态势分析模型迁移方法。一方面，为将比赛风格和队伍实力纳入对进球期望预测的考量，首先基于混合仿真数据集进行预训练，然后在特定仿真数据集上进行精调，生成一系列风格化态势分析模型；另一方面，为解决现实数据的小样本问题，首先在大规模仿真数据上进行预训练，然后在现实数据上进行精调。实验表明，本文的方法能够帮助态势分析模型：在仿真环境下，对原始的分析结果进行非线性矫正；在现实环境下，提升模型的性能表现。

（3）针对复杂场景下决策难以学习的问题，本文以足球决策智能体为研究对象，提出了一种基于态势分析辅助的决策学习方法。一是针对稀疏奖励问题，利用态势分析模型作为势能函数，设计了一种不改变智能体优化目标的引导性密集奖励；二是针对一致认知与个性涌现矛盾的问题，将态势分析网络模型的浅层进行迁移，设计了态势认知网络；三是针对样本分布不稳定带来的迁移困难问题，设计了两阶段的决策智能体训练算法，在多任务对抗学习阶段提取有关态势的通用深层表征，然后迁移至第二阶段强化学习训练过程，以解决不同任务、不同智能体间迁移知识的困难。实验表明，本文的方法极大地提升了强化学习智能体的训练速度以及训练收敛时的性能。

综上所述，本文以足球比赛为背景，深入研究了态势智能分析及决策辅助问题。本文的研究成果，一方面直接为足球比赛的推演、预测与复盘任务提供依据，另一方面有效辅助足球决策智能体的学习。

As one of the most popular sports in the world, football conveys to fans the cooperative spirit and tenacious will of athletes. The large space-time span, continuous and highly dynamic behavior, and multi-element coupling of football matches also pose numerous scientific challenges to researchers in various fields. Situational information expresses the current state and future development trends, which effectively assists decision-making. Hence, it has received extensive attention. Traditional situation analysis methods are mostly static statistical analysis for post-match replays, lacking dynamic analysis for confrontation deduction, so they are difficult to assist the real-time decision-making process. In recent years, the growth of football match data and the development of theoretical methods such as machine learning have provided new solutions. However, the low score, high uncertainty characteristics, and the difficulty of obtaining high-quality and large-scale data of football matches have brought great difficulties to the data-based football match situation analysis methods.

This dissertation takes football matches as the background and focuses on proposing an intelligent situation analysis method and utilizing situational analysis to assist football agents in decision-making learning. The main contributions and novelties of this dissertation are summarized as follows:

(1) To deal with the dynamic and uncertain challenges of football matches, expected goal based intelligent situation analysis method (All for Goals, AFG) for football matches based on large-scale data and deep learning is proposed. Firstly, the raw data is labeled by an automatic labeling method, and then a potential goal prediction model is trained based on deep learning. The concept of potential goal is extended to state evaluation, on-ball action evaluation, and off-ball running evaluation to quantify the relationship among any state, any player action, and potential goal probability. The experiment results indicate that, AFG achieves high recall rate in the goal prediction task; at the same time, the analysis results in real-time goal prediction, event credit assignment and off-ball running analysis tasks demonstrate consistency with domain knowledge.

(2) To deal with the fact that match style and team strength affect the situation analysis results and the difficulty of obtaining large-scale and high-quality football match data in the real world, a transfer method of the situation analysis model based on the pre-training and fine-tuning paradigm is proposed. On one hand, an pre-trained model is trained based on mixed simulation data set. Then fine-tuning is performed on the simulation data of specified style to introduce the match style and team strength into the consideration of potential goal prediction to form stylized situation analysis models. On the other hand, pre-training is performed on large-scale simulated data, then fine-tuning is performed on the real-world data to solve the small-sample problem. The experiment results indicate that, the method helps the situation analysis model: in the simulation environment, perform nonlinear correction on the original analysis results; in the real-world environment, improve the performance of model in the real-world.

(3) To deal with the difficulty of learning policy in complex scenarios, this dissertation takes football decision-making agents as the object and proposes a policy learning method based on situation analysis. Firstly, for the sparse rewards, the situation analysis model is used as a potential energy function to construct a guiding dense reward that does not change the optimization objective. Secondly, for the contradiction between consistent cognition and personality emerging, the shallow layers of the situation analysis network are transferred as the situation cognition network. Thirdly, for the transfer difficulties caused by unstable sample distribution in reinforcement learning, a two-stage training algorithm is proposed. In the first stage of adversarial multi-task learning, general situation deep features are extracted and transferred to the reinforcement learning in the second stage to solve the difficulty of transferring knowledge among different tasks and different agents in reinforcement learning. The experiment results indicate that, the method accelerates learning and improves the performance of agents when their learning converges.

In conclusion, this dissertation takes football matches as the background and studies intelligent analysis and decision-making assistance method of football match situation. The methods and results in this dissertation, on one hand, directly provide the basis for deduction, prediction, and replay tasks of football matches; on the other hand, effectively assist the learning of football decision-making agents.