Difference between revisions of "LightR"

Revision as of 07:14, 27 July 2022

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

Roadmap

One gun to rule them all

1. Learn a unique set of features per opponent, out of 100+ features

2. Learn a feature gate model for generalization to unseen bots

One movement to dodge everyone

1. Learn a model that uses hits and flattener waves simultaneously

2. Learn specialized movement patterns with reinforcement learning

3. Generalize to unseen bots with zero-shot & few-shot learning

@@ Line 9: / Line 9: @@
 ; Design principle
-: Strategy light, machine learning heavy
+: Strategy light, machine learning heavy.
-; Central goal
+; Roadmap
-: Learned models -> learned systems
+: One gun to rule them all
+:: 1. Learn a unique set of features per opponent, out of 100+ features
-; Planned experiments
+:: 2. Learn a feature gate model for generalization to unseen bots
-: Towards deep learning:
+: One movement to dodge everyone
-:: Multiple hand-tuned danger models -> Expert model & gate model
+:: 1. Learn a model that uses hits and flattener waves simultaneously
-:: Hand-crafted features with naive KNN -> Search-based sequence model
+:: 2. Learn specialized movement patterns with reinforcement learning
-:: Offline pre-training & online fine-tuning of everything above.
+:: 3. Generalize to unseen bots with zero-shot & few-shot learning
-: Towards differentiable programming:
-:: Directly optimizing prior probability of getting hit (max escape angle, distancing and multi-wave risk fusion)
-:: Per-instance level optimization of the above (Pareto frontier)
-:: List-wise modeling of targeting & surfing
 __NOTOC__ __NOEDITSECTION__
 {{Template:Bot Categorizer|author=Xor|isMega=true|isOneOnOne=true|isMelee=false|isOpenSource=false|extends=Interface}}