Match Tracing

Abstract

Win prediction and performance evaluation are two core subjects in the sport analytics. Traditionally, they are treated separately and studied by two independent communities. However, this is not the intuitive way how humans interpret the matches: we predict the match results with the competition carrying on, and simultaneously evaluate each action based on the game context and its downstream impact. Predicting the match outcomes and evaluating the actions are coupled tasks, and the more accurately we predict, the better the evaluation is. To this end, we develop a unified Match Tracing framework (namely, MT), for tackling the win prediction and performance evaluation jointly. The main idea of MT is to learn a real-time lookahead win rate curve rather than a single scalar (win or lose). And the value of an action can be objectively measured with respect to the increase or decrease of the curve. To meet the low-latency restrictions of the online deployments, an efficient model equipped with recurrent attention mechanism and matrix perturbation (i.e., MT-Net) is built for learning and yielding the win rate curve. MT-Net encodes the players' behavior sequence through an attention mechanism and captures the player-interaction effects through a graph embedding method. With the action values derived from the win rate curve, performance can be quantified at different granularities (action/player/match level) by integrated analysis. Experiments on an e-sport game demonstrate the prediction effectiveness and the feasibility of the MT framework. Furthermore, we present the detailed application cases of MT, including key actions recognition and close match detection.

Time to Die 2: Improved in-game death prediction in Dota 2

“…They predict events including item usage, fights, game end with 95% accuracy though this would still cause three mistakes per second when parsing livestream data at 60 FPS. Wang et al (2020) use a time-enabled transformer recurrent neural network to analyse behaviour sequences in the lead-up to important in-game events as described in the game log.…”

“…These approaches use data snapshots capturing only one, or occasionally, multiple time slices. More recent work aims to capture temporal patterns using neural network algorithms such as recurrent neural networks (RNNs) (Silva, Pappa, & Chaimowicz, 2018;Wang et al, 2020;White & Romano, 2020;Yu, Zhang, Chen, & Xie, 2018), long short-term memory NNs (Akhmedov & Phan, 2021) (including bidirectional LSTMs Kim & Lee, 2020) and two-stage spatial temporal networks (Yang et al, 2022). Wang et al (2020) use a time-enabled transformer recurrent neural network to encode behaviour sequences, capture important events and predict the winner.…”