search

kevanoullio / apples-to-apples-agent

Apples-to-Apples game with AI agent using various natural language processing and machine learning techniques.
GNU General Public License v3.0
0 stars 2 forks source link

Resulting "score" in model is the exact same in every round/game #40

Closed kevanoullio closed 2 months ago

kevanoullio commented 2 months ago

In regular game play mode. Shouldn't matter if losing/extra are selected.

For example, in the log file in debug mode, score: 2.1497700698184516 will be calculated 49 times in a 5 points to win 5 game session, and best score is updated 18 times with that value. It seems that the score is not significantly different and there are ties occurring and the AI as judge chooses the first red apple in the list, thus the first player initialized in the game will be chosen as the winner.

kevanoullio commented 2 months ago

it looks like the linear regression is outputting a very small target slope (not all the time, but seems to be when calculating the slopes to choose winning red apple), and then when calculating the score it's effectively like 2.14 - 0 or whatever it is and it seems to be the slope_predict that's the culprit, it seems to be at the beginning of the game or something, perhaps when there's no red/green apple pairs to make a prediction from:

slope_predict: [ 0.00000000e+00 1.13820910e-17 5.11324491e-17 -1.71796019e-16 8.20251909e-17 -7.40909594e-17 -5.22731752e-17 4.64700299e-18 1.13364512e-16 -3.80985046e-16 1.29753807e-16 -6.93946935e-17 0.00000000e+00 7.27858572e-17 0.00000000e+00 -6.69035656e-17 -6.76787849e-17 0.00000000e+00 0.00000000e+00 4.24865935e-16 [...]