AI Arena is a sophisticated game that integrates advanced machine learning techniques to train and optimize fighter models. Players can choose between different model types, each offering unique advantages and challenges. The game emphasizes strategic data collection and processing to enhance fighter performance, making it essential for players to understand and utilize the various settings and modes available.
We had the opportunity to sit down with the All-Star Invitational winner and G3 representative, Ixtlanian, for a full AI Arena guide on how to become a pro player and what are the essentials you need to nail down before you level up your game. This is the first guide in our series, and we will focus on introducing new players to the Advanced model.
Fighter Models
In AI Arena, there are three types of fighter models available:
- Simple - Ideal for beginners and quite popular but has limitations that prevent high-level competition.
- Original - Utilizes machine learning and sophisticated tools for fine-tuning, but gameplay can be very challenging. For example, demonstrating an action in one part of the map can unpredictably affect behavior elsewhere. This model is the least popular.
- Advanced - Combines the best features of the first two models, allowing for isolated zone edits without affecting the rest, and offers advanced settings to convert suboptimal data into useful training material. This model is favored by top players and will be the focus here.
Choosing the Advanced Model
Select the "Advanced" model before starting your training to utilize this guide. Enable the "Exponential Memory Updating" checkbox for more effective recent training sessions, which is suitable for significant changes to your fighter's behavior.
Data Collection Process
The game's official YouTube channel offers extensive videos on data collection. Here are some tips directly from Ixtlanian:
- Minimize Empty Information Slots: Shorten the pause between pressing the data collection button and performing an action to enhance training efficiency.
- Avoid Unnecessary Actions: Ensure actions are clear and isolated in each situation during data collection.
Data Processing
We will focus on advanced settings rather than simple ones. Refer to the tutorial in the upper left corner for detailed explanations of settings like “Epochs,” “Batch Size,” “Learning Rate,” and “Direction and Action Lambda.”
Tips for those familiar with the tutorial:
- Maximum Change Settings: Use maximum values for “Epochs” and “Learning Rate” combined with minimum values for “Batch Size” and “Direction and Action Lambda” for the initial training sessions to establish basic behavior.
- Protect Existing Data: Use maximum “Direction and Action Lambda” values to protect existing data, reducing “Epochs” and “Learning Rate” values for desired changes while maintaining current progress.
- Remove Sparsity: Enabling the “Remove Sparsity” checkbox removes most empty cells of information, making training more effective. This is used 95% of the time.
- Oversampling: Helps fighters distinguish and memorize different positions and directions, reacting more precisely to enemy actions. This mode is often used in initial training to lay the groundwork.
- Multi-Stream: Proportionally increases the chance of applying existing actions and directions if confirmed in current training. However, it may negatively affect other actions if intensive training for new actions is applied.
Buckets
Buckets are divided into two categories: “On Ground” and “In Air.” Actions shown in one category do not affect those in another unless selected. This isolation is a significant advantage of the advanced model, preferred by top players.
On Ground Buckets
- Projectile Active: Activated only when special abilities are used and only while your fighter is on the ground.
- Opponent Stunned and Opponent Knockback: Triggered when the opponent is stunned or knocked back, respectively.
- Opponent Active: Used in all other cases when your fighter is on the ground, including all options until the opponent is stunned or knocked back.
In Air Buckets
- Near Blast Zone, Safe Zone, Under Stage, Side Of Stage: Divide the airspace into zones where your fighter is located. These zones are trained in isolation from each other.
Focus Settings in AI Arena
After selecting the BUCKETS in which to make changes, we move on to what our fighter should emphasize. Most of them are also described in the Tutorial inside the game, but we will focus on the most frequently used ones and give hints on how they work in different combinations.
- Your Actions - This setting memorizes your fighter's previous actions and allows you to build a sequence of actions. For example, you can train your fighter to execute a head kick after an uppercut. It is rarely used, mostly in final training sessions.
- Opponent Actions - This setting allows your fighter to link their actions to the opponent's actions. For example, your fighter can put up a shield when the enemy attacks or grab the enemy when they are shielded. It is almost always used in ground training from the very beginning.
- Your Positioning - This helps your fighter understand their location in space to use the correct directions or actions. It is always used to return to the platform from "Side Of Stage" and "Near Blast Zone" and is rarely used on the ground.
- Opponent Positioning - This allows your fighter to understand the opponent's location in space. It is rarely used in the current meta but can be experimented with.
- Your Direction - This considers where your fighter is looking in relation to the enemy. It is almost always used on the ground and in the “Safe Zone” to chase the enemy and determine if the fighter is facing the enemy or has their back to them. It is often used in conjunction with "Angle to Opponent" and is not used for returning to the platform from "Side Of Stage" and "Near Blast Zone".Hint: Use this to grab an enemy only when you're facing them or to do an uppercut.
- Opponent Direction - This considers where the enemy is looking, such as to block only when they attack in your direction. It is used in combination with "Angle to Opponent".
- Angle to Opponent - This determines your fighter's position relative to the enemy and is used whenever actions are oriented toward the enemy, most often in conjunction with "Your Direction".
- Recoveries Remaining - This setting is used only to return from "Side Of Stage" and "Near Blast Zone." It helps your fighter understand which techniques they have already used to return, increasing the likelihood of using other techniques for a successful return.
- Discrete Distance - This helps your fighter understand whether they are close or far from the enemy, crucial for vertical combat in the air and all ground combat. It is always used in training.
- Proximity to Edge - This helps your fighter understand whether they are on the edge of the platform or in the middle, introducing variety in behavior. It is very important to master and is used only while on the ground, often in combination with "Discrete Distance," "Angle to Opponent," and "Your Direction."
- Opponent Off Stage - This allows your fighter to realize when the enemy is outside the platform to effectively defend the edge and prevent the enemy's return. It is very important to master and is used only while your fighter is on the ground, often in combination with "Discrete Distance," "Angle to Opponent," and "Your Direction."
Pro Tips from Ixtlanian
To create an adaptive and unpredictable fighter, use various focus combinations to teach specific behaviors in response to different opponent actions. This approach ensures that your fighter can quickly and accurately react to the opponent's moves, avoiding predictability and becoming more versatile in combat situations.
Carefully plan your training sessions by determining the specific skills you want your fighter to learn. Demonstrate these actions during training, and process the data with settings that will optimize their effectiveness. For example, if you aim to teach your fighter to perform an uppercut or flipkick in the air at close range and to jump closer for a punch at long range, you should select "Your Direction," "Angle to Opponent," and "Discrete Distance." These settings help your fighter understand the direction of their gaze, their angle relative to the enemy, and the distance between them.
Finally, always experiment and iterate. If the results of a training session are unsatisfactory, do not save the data. Instead, identify what went wrong, make necessary adjustments to the training or data processing, and try again. This iterative process will enhance your training sessions over time, resulting in a fighter with superior reaction time and relevance.