Support NeoGAF

[cormack12]

I love this guys channel.



Chapters
[00:00] Intro
[00:50] What is a Drivatar?
[02:03] The Challenge of Racing AI
[05:22] Forza Physics
[09:01] The Original Drivatar
[13:30] Drivatar v2.0
[18:21] Designer Modifications
[20:33] Hacks & Tweaks
[22:27] Closing


Summary

• Was founded at Microsoft Research, Cambridge, UK;
• One of the longest running applications of machine learning in the game industry;
• Based on player modelling, if we can gather enough data about how players play then an AI system can begin to make decisions based on that knowledge;
• What makes drivatars so special is that it has been in the Forza series since the very beginning, meaning its development started in the first entry of Forza Motorsport (2005);
• It's impressive because it does more than just replay actions the player has done before in the same circumstance;
• It makes intelligent decisions based on our data in situations you may never have experiencd before (e.g. unknown track, unknown car);
• This is where machine learning rather than sim modelling AI really shines;
• AI is usually full of logic and choices that tell the AI what to do in any given situation;
• It's increasingly difficult and time consuming to do this given the scope and scale of a lot of modern games;
• Machine learning is able to take all the data, compress it into patterns and apply generalised rules to those patterns;
• Underpinned by the excellent and accurate full physics system;
• Despite the game running at 30/60 fps the physics system runs at 360fps;
• The physics systems are layered and abstracted to allow the split between Horizon and Motorsport;
• The race lines we see are generated by the AI controller. These give us the optimum path through the course given the current utility;
• Drivatars are essentially small pieces of data called artificial neural networks;
• While the neural network itself will be a consistent logical topology, the modifications are managed by the weights on each connection between them;
• To get undefined outputs requires the use of Bayesian artificial neural networks;
• Drivatars allows these optional paths to have probability applied to them. This means it would be more probable that the drivatar would ease up on gas then accelerate to get a wider corner, than brake and get back on the perfect line. Similarly, it might be more probable the player slams on and then understeers the corner;
• New system is not properly public but creative director has said it uses 'deep learning';
• Drivatar 2.0 tracks 12 different type of segments/turns;
• It also constantly tracks the player/drivatar accuracy and will adapt as the player improves;
• Rewind feature is ignored to prevent 'gaming' of the system and creating perfect AI;
• Some behaviours are tweaked to be disallowed (e.g. donutting, driving backwards, smashing into players on corners as barrier brakes);
• Drivatars do not rubber band AI based on the leader, but they do have vehicle rubber banding (less performant, weigh more to slow acceleration, less torque available or scaling back tyre friction). If you're losing, this scaling is applied backwards so your car becomes more performant and has better utility;
• All this is tied to the full physics model of the cars so it doesn't break the game though, it is the optimum the vehicle is capable of or the worst given wear, and misconfiguration;
• There is one driver in the game that isn't a drivatar which is 'The Stig' which is a manually tuned AI;

 

[M1chl]

I love this guys channel.



Chapters
[00:00] Intro
[00:50] What is a Drivatar?
[02:03] The Challenge of Racing AI
[05:22] Forza Physics
[09:01] The Original Drivatar
[13:30] Drivatar v2.0
[18:21] Designer Modifications
[20:33] Hacks & Tweaks
[22:27] Closing


Summary

• Was founded at Microsoft Research, Cambridge, UK;
• One of the longest running applications of machine learning in the game industry;
• Based on player modelling, if we can gather enough data about how players play then an AI system can begind to mak decisions based on that knowledge;
• What makes drivatars so special is that it has been in the Forza series since the very beginning, meaning its development started in the first entry of Forza Motorsport (2005);
• It's impressive because it does more than just replay actions the player has done before in the same circumstance;
• It makes intelligent decisions based on our data in situations you may nevr have experiencd before (e.g. unknown track, unknown car);
• This is where machine learnign rather than sim modelling AI really shines;
• AI is usually full of logic and choices that tell the AI what to do in any given situation;
• It's increasingly difficult and time consuming to do this given the scope and scale of a lot of modern games;
• Machine learning is able to take all the data, compress it into patterns and apply generalised rules to those patterns;
• Underpinned by the excellent and accurate full physics system;
• Despitee the game running at 30/60 fps the physics system runs at 360fps;
• The physics systems are layered and abstracted to allow the split between Horizon and Motorsport;
• The race lines we see are generated by the AI controller. These give us the optimum path through the course given the current utility;
• Drivatars are essentially small pieces of data called artifical neural networks;
• While the neural network itself will be a consistent logical topology, the modifications are managed by the weights on each connection between them;
• To get undefined outputs requires the use of Bayesian artificial neural networks;
• Drivatars allows these optional paths to have probability applied to them. This means it would be more probable that the drivatar would ease up on gas then accelerate to get a wider corner, than brake and get back on the perfect line. Similarly, it might be more probable the player slams on and then understeers the corner;
• New system is not properly public but creative director has said it uses 'deep learning';
• Drivatar 2.0 tracks 12 different type of segments/turns;
• It also constantly tracks the player/drivatar accuracy and will adapt as the player improves;
• Rewind feature is ignored to prevent 'gaming' of the system and creating perfect AI;
• Some behaviours are tweaked to be disallowed (e.g. donutting, driving backwards, smashing into players on corners as barrier brakes);
• Drivatars do not rubber band AI based on the leader, but they do have vehicle rubber banding (less performant, weigh more to slow acceleration, less torque available or scaling back tyre friction. If you're losing this scaling is applies backwards so your car becomes more performant and has better utility;
• All this is tied to the full physics model of the cars so it doesn't break the game though, it is the optimum the vehicle is capable of or the worst given wear, and misconfiguration;
• There is one driver in the game that isn't a drivatar which is 'The Stig' which is a manually tuned AI;

For me this is one of the most fascinating. tech in the industry. I also sometimes do some machine learning. so since then my interest in Drivatars rised.

 

[wolfboy203]

Drivatars seem useless when they still give the top a.i. racers advantages like extra horsepower, unlimited grip, unrealistic speed boosts, etc

 

[Fox Mulder]

The tech sounds neat, but it’s kind of dumb when someone I know that has played the game for like 5 minutes on my friends list is loaded in a race with me. I know they don’t have shit for data on them.

plus they still cheat.

 

[MAX PAYMENT]

This stuff is super interesting. I had fun reading up on how an AI was developed for a board game. It's basically a series of "best practice" solutions for common problems. Fascinating field.