DSC Conference Day 1
Clustering Players’ Strategies for Generation of NPCs, by Pandora
The game in focus is Mythic Legends. It features an asynchronous multiplayer system that matches players against “snapshots” of other players. This system eliminates the need for both players to be online simultaneously and can handle situations where a player disconnects.
TL;DR
This is a showcase of how a problem, such as the need to create a new event quickly, can lead to innovative solutions. In this case, the solution was to use data to create custom AI teams for players to compete against in a game that is primarily based on PVP.
Presentation
The main problem was the creation of daily events. These events needed to feature progressively harder battles. However, the use of snapshots was not possible as they only represented single heroes, not teams. New encounters and boss fights required templates. The events needed to be more fun, unusual, and serve as an extended tutorial. Additionally, these events were also used for testing new classes.
The challenge was to create at least 1000 templates. The question was, how could this be achieved? Manual creation by a game designer was too much work. Using the most frequent snapshots was not a viable option as the resulting templates would have a low skill level and would not be challenging for players. Using the top win rate was also not feasible as these formations are only effective against specific formations. The solution was found in snapshot clusters. By grouping similar hero races, diverse classes and an overall skill level could be achieved.
To compare differences between teams, several factors were considered: stat comparison, placement comparison, and cluster comparison (composition of team).
The success of this approach is measured by the number of players sticking to the event. This may be biased as some players may only be interested in the reward. However, if players improve their win rate after the event, it indicates success. Additionally, this approach saves time for game designers.
Conclusion
To achieve success, it’s important to understand the product and its goals. Reusing knowledge, concepts, and code can be beneficial. These elements need to be translated into data science tasks. Whenever possible, simplify the process. Make use of available data and apply domain knowledge and manual work where necessary.
Stories about Industrializing In-Game ML Solutions
This section presents stories from Pavle Petrovic, the technical lead of Ubisoft Belgrade, a subdivision of the Technology Group in Ubisoft. He shares insights into how they are implementing Machine Learning (ML) tools on the developer side and their plans to implement these tools on the player side as well.
Presentation
In the early stages of ML implementation at Ubisoft, all ML operations were handled by a separate Python code stack. This setup required the game agent to constantly wait for the Python stack to complete its tasks, which significantly slowed down the process. This arrangement also complicated the prototyping and porting of games.
Currently, only the training process occurs outside of the game engine, as the entire ML framework is now integrated within the game engine itself. While this new architecture has improved game speed, it has also introduced a new challenge: data scientists are unable to directly access data, which slows down the entire iteration process.
The ML tools are utilized in various aspects of the game, including auto vehicle driving, character map navigation, and bot fights in “For Honor”. The development process involves data scientists creating tools, the tech group providing these tools, and gameplay programmers implementing the recipes.
One example of the use of these tools is in imitation learning for aiming. This process involves training from player data, using the ONNX runtime inference engine, and an offline model with an imitation learning algorithm. The agent focuses on dangerous points in a shooter game, such as doorways, corners, and staircases.
Another application is in navigation, where online Reinforcement Learning (RL) is used, with data and models generated during training.
An interesting challenge arose in “Rainbow 6 Siege”, where the players were too skilled, the bots were too weak, and new players struggled to learn the game. The solution was the introduction of a VS AI mode, which uses imitation learning for aiming to help new players compete against better AI.
Autonomous Driving - Breakout Session
This session was presented by HTEC and Rivian.
Introduction
The term “autonomous” refers to fully autonomous vehicles that operate without any human interactions. This does not include robot taxis, for example. The Society of Automotive Engineers (SAE) has defined levels of driving automation from L0 to L5.
L0 represents vehicles with blindspot detection systems that only provide warnings. L1 includes ADAS systems that control accelerating, braking, and emergency braking. L2 represents vehicles that have two or more functions automated, such as emergency obstacle avoidance.
L3, which was approved this year, represents the highest level of automation currently commercially available. At this level, drivers should be able to take their hands off the wheel. L4 allows drivers to take their eyes off the road to do things like watch videos or use their phone. L5, the highest level, allows passengers to simply sit down and let the vehicle drive.
BMW has reached commercial L3, while robo taxis are aiming for L5. These are the two main focuses of the industry at the moment.
Functional safety and cybersecurity standards require extensive testing before being implemented in production. Thousands of vehicles need to run without error for decades before achieving commercial L5.
Improving the quality of autonomous driving systems will require collecting more and more data. For example, encountering a truck that’s tipped over on its side on the highway is a situation that current systems may not know how to react to, or not giving a firetruck priority.
Certification of autonomous vehicles often happens in places with great weather, which doesn’t take into account varying road conditions.
Data Analytics in Gaming
This section is presented by Siyu Sun from Gameloft Paris, who is also an ex-Blizzard employee.
Data science has various applications in gaming. These include game analytics, which can help identify points where players get stuck, marketing analytics, and the use of Data Science (DS), Machine Learning (ML), and Artificial Intelligence (AI) to capture market trends.
There are different levels of data science used in gaming. Predictive analysis, for instance, aids in decision-making. Predictions can also be used as a product or system, such as an in-game system for players or for marketing purposes. AI is also used by gaming developers.
Monetization in Free-to-Play (F2P) Games and Their Optimization
Monetization in F2P games can be achieved through various methods. In-app purchases (IAP) are one such method. The primary goal of IAP is to improve the experience for both free and paying players. It involves predicting if people will pay or buy, analyzing the conversion points (when it switches from being doable for free to being extremely hard to play without paying), and understanding what types of items are bought (bundles, on sale, etc.). The main challenges are ensuring fairness (is it fair to give personally adapted prices to players?) and ethical ML (AI cannot be the only deciders in what to offer players). The solutions include giving choices to players (opt-in, opt-out) and maintaining transparent communication (when promotions happen, etc.).
Ads are another method of monetization. There are many different types of ads, and the key question is how often they should be displayed. Too many ads can lead to a bad experience, so a balance is needed.
Subscriptions, such as battle passes, offer a premium experience and have been a great innovation since 2021. Implementing tiers of subscriptions and balancing pricing vs rewards (each season needs to be comparable) are important aspects of this method. A big bonus is that it “forces” engagement to get all rewards. However, it is hard to test due to its duration.
Pay-to-win is another method of monetization, albeit a controversial one.
A/B testing, which involves having two different versions of the game and seeing which one works better, is the golden standard of testing.