Название: Game AI Uncovered: Volume Two
Автор: Paul Roberts
Издательство: CRC Press
Год: 2024
Страниц: 227
Язык: английский
Формат: pdf (true)
Размер: 15.6 MB

Game AI Uncovered: Volume Two continues the series with the collected wisdom, ideas, tricks and cutting‑edge techniques from 22 of the top game AI professionals and researchers from around the world.

The techniques discussed in these pages cover the underlying development of a wide array of published titles, including The Survivalists, Wheelman, Plants vs. Zombies: Battle for Neighborville, Dead Space, Zombie Army 4, Evil Genius 2, Sniper Elite 5, Sonic & All‑Stars Racing Transformed, DiRT: Showdown, and more.

Contained within this volume are overviews and insights covering a host of different areas within game AI, including generalised planners, player imitation, awareness, dynamic behaviour trees, decision‑making architectures, agent learning for automated playthroughs, utility systems, Machine Learning for cinematography, directed acyclic graphs, environment steering, difficulty scenarios, environmental cues through voxels, automated testing approaches, dumbing down your AI, synchronized path following, and much more.

The graph frames are stored in the OpenTimelineIO format, where each shot is stored as a separate clip and each clip has each frame within it stored as metadata to the clip in NetworkX format, stored as a JSON object. A clip‑per‑shot is stored using the shot boundaries previously determined using the Movienet pipeline tools. This relies on the PyShotDetect library to detect shot boundaries, but the essence is that a number of algorithms are used to determine changes in camera/lighting/scene composition. An OpenTimelineIO ‘.otio’ file is constructed such that each clip in the timeline represents a shot detection boundary, this clip is then used to process the contents for the shot. The initial frame of each shot is determined to be a keyframe and a NetworkX graph is stored as metadata for the clip. This graph is stored as JSON format data within the metadata of the ‘.otio’ file and can be parsed by the NetworkX Python graph reading library.

As games are growing ever larger and more complex, the need for testing, and the time it takes to test a game, has increased. There are many well‑established types of automated tests like unit tests, functional tests, and smoke tests. These tests gather performance metrics like frames per second, memory usage, loading times, bandwidth usage, and many more. They also perform code coverage, detect crashes, fire asserts, and create error logs, amongst a plethora of other functions. This data is invaluable for development teams to optimise and stabilise their games.

This chapter will focus on automated testing i.e., tests that play the game as a player by simulating keyboard, mouse, and game controller inputs. Most solutions that are used to test games like this use a combination of pre‑generated scripts, behaviour trees (BTs), finite‑state machines (FSM), and reinforcement learning agents. They load a level and test it for a certain amount of time and then quit. Taking this a step further, an AI framework will be described that is capable of consistently completing entire games from start to finish. The benefits of such a system are manyfold. Not only does it provide more extensive code coverage and metrics that better reflect the game, but it can also detect game and level design issues. If the framework is unable to complete a game, then this is usually due to a coding bug or issue with the level or game design.

Beginners to the area of game AI, along with professional developers, will find a wealth of knowledge that will not only help in the development of your own games but also spark ideas for new approaches.




ОТСУТСТВУЕТ ССЫЛКА/ НЕ РАБОЧАЯ ССЫЛКА ЕСТЬ РЕШЕНИЕ, ПИШИМ СЮДА!