Skip to main content

LUX S2 Submission Analysis

Created on April 25|Last edited on April 30

Submission Run


5M10M15M20Mglobal_step200400600800
5M10M15M20Mglobal_step010000200003000040000
Run set
3




A disappointing follow-up run

After the excellent 1c3f35f4 run, I had hopes that I could get excellent follow-up results:
b4b9810 close_extras takes arguments that have to be handled
66ed1e5 Small and medium use same map size for eval
63681a9 Fix to support different eval map vs train map
2c371c1 Fix imports
2684f03 mean score_function for eval
5f3766b Move Lux StatsTracking into a lux/stats module
4f3b153 Eval only tracks win/loss
11149af factories_alive stat awards current number of factories
bbfa3ee medium-debug env hyperparams for local
133c8eb medium 30M steps (should complete in 1 day)
f9d6769 LuxAI_S2-v0-medium nearly guarantees existence of ice
d70c4ac Tweak small to not print as much and debug
70c8be3 LuxAI_S2-v0-small is a 16x16, 1 factory map
The only apparent difference in behavior is the FACTORIES_ALIVE reward weighed at 0.02