Shovas
Premium
I would like to see a streamlined user experience for automatically adjusting the AI level based on real driving performance.
The problem is wasted time. It takes so long to find the right AI level that it takes away from the enjoyment of the experience. I find this to be the case in GTR2, rFactor2, AMS2 and I'm not aware of a sim that has improved in this area. RaceRoom has tried but I don't think it really works the way it needs to (see below).
The goal is to ensure that the player feels confident that one practice and one qualifying will produce a satisfyingly competitive race.
The only system like this that I'm aware of that even tries to tackle this is RaceRoom's Adaptive AI and it takes many wasted races to adjust. Furthermore, track familiarity doesn't transfer so after running a race at one track, that adjusted AI level doesn't necessarily apply to a new track. It also doesn't take into account increasing skill with the simulator or decreasing skill (ex. you've not raced in a month).
I think the best inspiration is from real heat-based systems like iRacing where qualifying acts as a heat from which the system can group racers by their *real, current performance* into splits.
What makes iRacing's multiplayer work is the sheer number of racers in any given session: There's enough to have the fine-grained performance levels that can be split and each split has a relatively competitive race.
This sheer number of drivers is the key to an automatically adjusting AI difficulty level. It's not enough to have one AI skill profile per car (ie. AI Talent files in rFactor-lineage games). It's not enough to know their relative performance based on skill profiles. You must actually know their *real performance* on the track in the session with the variables that are presented, such as temperature, humidity, wind, etc.
In a Big Data machine learning system, you could actually simulate all AI in all kinds of scenarios and that would probably work well.
You could also crowdsource real driver laps from real players playing the game, even after the game is released. I believe this is how Forza Horizon does it. That could make it an automatic and self-improving system.
The point is to race and practice and qualifying are just a means to get there. If practice and qualifying don't guarantee a satisfyingly competitive race performance by the AI then you've wasted upwards of an hour or more. But if practice and qualifying are representative of race peformance then there's nothing left to do but enjoy great racing.
Here's hoping!
The problem is wasted time. It takes so long to find the right AI level that it takes away from the enjoyment of the experience. I find this to be the case in GTR2, rFactor2, AMS2 and I'm not aware of a sim that has improved in this area. RaceRoom has tried but I don't think it really works the way it needs to (see below).
The goal is to ensure that the player feels confident that one practice and one qualifying will produce a satisfyingly competitive race.
The only system like this that I'm aware of that even tries to tackle this is RaceRoom's Adaptive AI and it takes many wasted races to adjust. Furthermore, track familiarity doesn't transfer so after running a race at one track, that adjusted AI level doesn't necessarily apply to a new track. It also doesn't take into account increasing skill with the simulator or decreasing skill (ex. you've not raced in a month).
I think the best inspiration is from real heat-based systems like iRacing where qualifying acts as a heat from which the system can group racers by their *real, current performance* into splits.
What makes iRacing's multiplayer work is the sheer number of racers in any given session: There's enough to have the fine-grained performance levels that can be split and each split has a relatively competitive race.
This sheer number of drivers is the key to an automatically adjusting AI difficulty level. It's not enough to have one AI skill profile per car (ie. AI Talent files in rFactor-lineage games). It's not enough to know their relative performance based on skill profiles. You must actually know their *real performance* on the track in the session with the variables that are presented, such as temperature, humidity, wind, etc.
In a Big Data machine learning system, you could actually simulate all AI in all kinds of scenarios and that would probably work well.
You could also crowdsource real driver laps from real players playing the game, even after the game is released. I believe this is how Forza Horizon does it. That could make it an automatic and self-improving system.
The point is to race and practice and qualifying are just a means to get there. If practice and qualifying don't guarantee a satisfyingly competitive race performance by the AI then you've wasted upwards of an hour or more. But if practice and qualifying are representative of race peformance then there's nothing left to do but enjoy great racing.
Here's hoping!
Last edited:
