Learning the patterns of balance in a multi-player shooter game

Abstract

A particular challenge of the game design process is when the designer is requested to orchestrate dissimilar elements of games such as visuals, audio, narrative and rules to achieve a speci c play experience. Within the domain of adversarial rst person shooter games, for instance, a designer must be able to comprehend the di erences between the weapons available in the game, and appropriately cra a game level to take advantage of strengths and weaknesses of those weapons. As an initial study towards computationally orchestrating dissimilar content generators in games, this paper presents a computational model which can classify a matchup of a team-based shooter game as balanced or as favoring one or the other team. e computational model uses convolutional neural networks to learn how game balance is a ected by the level, represented as an image, and each team's weapon parameters. e model was trained on a corpus of over 50,000 simulated games with arti cial agents on a diverse set of levels created by 39 di erent generators. e results show that the fusion of levels, when processed by a convolutional neural network, and weapon parameters yields an accuracy far above the baseline but also improves accuracy compared to articial neural networks or models which use partial information, such as only the weapon or only the level as input.

Pairing character classes in a deathmatch shooter game via a deep-learning surrogate model

“…2a that the score of player 1 is spread over the entire spectrum, from extreme advantage over player 2 (scores near 1.0) to extreme disadvantage (scores near 0.0) and a balanced match (score of 0.5). Unlike [17], the score does not follow a normal distribution centered around 0.5; the distribution is rather uniform, with occurrences decreasing towards the edge cases. This points to a rich dataset with positive and negative examples of a balanced match.…”

“…considering only the structural parts of the level [22] or the properties of weapons [10] in a vacuum. Compared to earlier work [17], the model introduced in this paper uses a much broader corpus of structurally and ludically complex levels and a diverse set of classes as its training set. It is also the first instance that uses such a multi-faceted computational model as a surrogate in generating aspects of a game's ruleset for specific levels and intended gameplay outcomes, while in [18] a similar surrogate model is used to drive the evolutionary adaptation of levels in order to balance a specific set of character classes.…”

“…Based on [17,18] and additional preliminary experiments with different network architectures, learning rates and activation functions, the best performing CNN architecture used for reported experiments has two separate branches: one for level structure and one for class parameters. The character class branch consists of a single fully-connected layer of 8 nodes that transforms the 16 character class parameters (8 per player) into 8 features.…”

Orchestrating Game Generation

“…Given sufficient data, we can assume that a deep learning approach can assess whether game content match. For instance, deep learning was used to predict how levels and weapon parameters, combined, affect gameplay balance [109]; in [110], a similar mapping was used to adapt hand-authored or generated levels to be more balanced for specific matchups between character classes. It has been already shown that machine learning can capture level structures from gameplay videos [52], which can drive the generation of new levels.…”

A literature review of Empirical Evidence on Procedural Content Generation in Game-Related Implementation

“…Based on the current state of digital games and its nature, as well as unfathomable speed of how computation power is being developed, the development of game contents is moving at similar pace as well [8]. Conventional content generation in digital games is a very rigorous and time-consuming process [9], much of its development pipeline requires multiple parts of development lifecycle and requires multiple expert to validate its output [10].…”