Difference between revisions of "ScalarR/Random thoughts"
< ScalarR
Jump to navigation
Jump to search
(Some random thoughts) |
m |
||
| (3 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
| + | This page is a collection of some random thoughts. Feel free to start a discussion. | ||
| + | |||
| + | |||
| + | === On making a top bot === | ||
| + | |||
; What really makes a bot strong overall? | ; What really makes a bot strong overall? | ||
| − | : Excessive tuning of every single detail. See the logs in [[ScalarR/Version History|/Version History]]. | + | : Excessive tuning of every single detail (against the entire Rumble). See the logs in [[ScalarR/Version History|/Version History]]. |
; What affects performance the most? | ; What affects performance the most? | ||
| Line 6: | Line 11: | ||
; How to make a top movement from scratch? | ; How to make a top movement from scratch? | ||
| − | : First make it right. Then spend some time tuning bandwidth and distancing. Good aiming models can be used as good danger models, further tuning may not help much. Flatteners aren't necessary. | + | : First make it right. Then spend some time tuning bandwidth and distancing. Good aiming models can be used as good danger models, further tuning may not help much. Flatteners aren't necessary. (ScalarR managed to get 91 APS and 99.91 PWIN without a flattener.) |
; How to make a top gun from scratch? | ; How to make a top gun from scratch? | ||
| − | : Basic attributes + wall attributes + machine learning on whole rumble data = good to go. Fancy features may help against | + | : Basic attributes + wall attributes + machine learning on whole rumble data = good to go. Fancy features may help against a couple of bots, but may not help a lot overall. |
; What may be the next big innovation? | ; What may be the next big innovation? | ||
: Active bullet shielding may help against the top bots, but about general rumble, there's still much room in accurately predicting & controlling the opponents' targeting. Reducing b in a / (a + b) is more efficient once a is larger. | : Active bullet shielding may help against the top bots, but about general rumble, there's still much room in accurately predicting & controlling the opponents' targeting. Reducing b in a / (a + b) is more efficient once a is larger. | ||
| + | |||
| + | |||
| + | === On why things work === | ||
| + | |||
| + | ; What makes wave surfing effective? | ||
| + | : Wave surfing is effective because it actively find movement patterns that the opponent is hard to hit, and keep the current pattern as long as no more hits are made. This explains why adding even a little flattening to even some simple learners hurts performance a lot, so does adding any randomness. This contradicts with the common belief that wave surfing does well by flattening the movement profile. Flattening the movement profiles is necessary, but not enough. | ||
| + | |||
| + | ; What makes anti-surfing effective? | ||
| + | : I don't know. Many anti-surfing guns learn from recent firing scans mainly, however BeepBoop outperformed all of them with mostly wall attributes. I would say there are still plenty of room to explore in this aspect, feel free to share your insights. | ||
Latest revision as of 07:46, 18 January 2023
This page is a collection of some random thoughts. Feel free to start a discussion.
On making a top bot
- What really makes a bot strong overall?
- Excessive tuning of every single detail (against the entire Rumble). See the logs in /Version History.
- What affects performance the most?
- Targeting and movement are equally important. Strong targeting helps more against strong opponents, while strong movement earns more from the weak.
- How to make a top movement from scratch?
- First make it right. Then spend some time tuning bandwidth and distancing. Good aiming models can be used as good danger models, further tuning may not help much. Flatteners aren't necessary. (ScalarR managed to get 91 APS and 99.91 PWIN without a flattener.)
- How to make a top gun from scratch?
- Basic attributes + wall attributes + machine learning on whole rumble data = good to go. Fancy features may help against a couple of bots, but may not help a lot overall.
- What may be the next big innovation?
- Active bullet shielding may help against the top bots, but about general rumble, there's still much room in accurately predicting & controlling the opponents' targeting. Reducing b in a / (a + b) is more efficient once a is larger.
On why things work
- What makes wave surfing effective?
- Wave surfing is effective because it actively find movement patterns that the opponent is hard to hit, and keep the current pattern as long as no more hits are made. This explains why adding even a little flattening to even some simple learners hurts performance a lot, so does adding any randomness. This contradicts with the common belief that wave surfing does well by flattening the movement profile. Flattening the movement profiles is necessary, but not enough.
- What makes anti-surfing effective?
- I don't know. Many anti-surfing guns learn from recent firing scans mainly, however BeepBoop outperformed all of them with mostly wall attributes. I would say there are still plenty of room to explore in this aspect, feel free to share your insights.
