Updates on Continuous Model Improvement for Fully-supervised State Prediction Model
#2
by
BXYMartin
- opened
This version is a Proof-of-Concept for the network to show if the capacity is enough to learn the mechanism behind the game. Fully-supervised State Prediction Model takes state and action (based on this state) as input and predict the next state.
aeb3b3379266a96f1f381162932df92bdf87c429
was overfitted on a small subset of tags and entities excluding the game and hero entity, namely:
GAME_TAG.ENTITY_ID,
GAME_TAG.HEALTH,
GAME_TAG.ATK,
GAME_TAG.COST,
GAME_TAG.ARMOR,
GAME_TAG.ZONE,
GAME_TAG.CONTROLLER,
GAME_TAG.CLASS,
GAME_TAG.CARDRACE,
GAME_TAG.CARDTYPE,
GAME_TAG.SECRET,
GAME_TAG.ZONE_POSITION,
So it means that the relationship between entity ids may not generalize well
38c184baddc99c943023eb62150180e554a052b2
included game entity, a wider range of tags including:
GAME_TAG.ENTITY_ID,
GAME_TAG.HEALTH,
GAME_TAG.DAMAGE,
GAME_TAG.ATK,
GAME_TAG.COST,
GAME_TAG.ARMOR,
GAME_TAG.ZONE,
GAME_TAG.CONTROLLER,
GAME_TAG.CLASS,
GAME_TAG.CARDRACE,
GAME_TAG.CARDTYPE,
GAME_TAG.SECRET,
GAME_TAG.ZONE_POSITION,
GAME_TAG.FATIGUE,
GAME_TAG.EXHAUSTED,
# Extra
GAME_TAG.DURABILITY,
GAME_TAG.ENRAGED,
GAME_TAG.FROZEN,
GAME_TAG.SILENCED,
GAME_TAG.WINDFURY,
GAME_TAG.TAUNT,
GAME_TAG.STEALTH,
GAME_TAG.POISONOUS,
GAME_TAG.CHARGE,
GAME_TAG.LIFESTEAL,
GAME_TAG.RUSH,
GAME_TAG.ECHO,
GAME_TAG.REBORN,
GAME_TAG.DEATHRATTLE,
GAME_TAG.COMBO,
GAME_TAG.IMMUNE,
GAME_TAG.MAGNET,
GAME_TAG.OVERKILL,
GAME_TAG.INSPIRE,
GAME_TAG.OUTCAST,
GAME_TAG.SPELLBURST,
GAME_TAG.DORMANT,
GAME_TAG.FRENZY,
GAME_TAG.TRADEABLE,
And trained on the difference between input state and expected output state to shorten the predicted outputs.