Difference between revisions of "LightR"

From Robowiki
Jump to navigation Jump to search
m
m (Towards differentiable programming)
Line 12: Line 12:
  
 
; Planned experiments
 
; Planned experiments
: Hand-made risk of distancing & hit -> Pareto-based multi-objective risk
+
: Towards deep learning:
: Multiple hand-tuned danger models -> Expert model & gate model
+
:: Multiple hand-tuned danger models -> Expert model & gate model
: Hand-crafted features with naive KNN -> Search-based sequence model
+
:: Hand-crafted features with naive KNN -> Search-based sequence model
: Slow networks -> Knowedge distill & quantization aware training
+
:: Slow networks -> Knowledge distill & quantization aware training
: Offline pretraining & online fine-tuning of everything above.  
+
:: Offline pre-training & online fine-tuning of everything above.  
 
+
: Towards differentiable programming:
 +
:: Directly optimizing max escape angle & prior probability of getting hit (distancing)
 +
:: Per-instance level optimization of the above (mea & distancing as part of network)
  
 
__NOTOC__ __NOEDITSECTION__
 
__NOTOC__ __NOEDITSECTION__
  
 
{{Template:Bot Categorizer|author=Xor|isMega=true|isOneOnOne=true|isMelee=false|isOpenSource=false|extends=Interface}}
 
{{Template:Bot Categorizer|author=Xor|isMega=true|isOneOnOne=true|isMelee=false|isOpenSource=false|extends=Interface}}

Revision as of 09:12, 19 March 2022

LightR Sub-pages:
Version History

This page is under construction. For recent activities, see Version History.


Design principle
Strategy light, machine learning heavy.
Planned experiments
Towards deep learning:
Multiple hand-tuned danger models -> Expert model & gate model
Hand-crafted features with naive KNN -> Search-based sequence model
Slow networks -> Knowledge distill & quantization aware training
Offline pre-training & online fine-tuning of everything above.
Towards differentiable programming:
Directly optimizing max escape angle & prior probability of getting hit (distancing)
Per-instance level optimization of the above (mea & distancing as part of network)