Predict next breed of an agent

Vayel commented 6 years ago

Using AI

Vayel commented 6 years ago

Train the network each time a request is made on the web UI

Vayel commented 6 years ago

Can use two models:

Naive: a generic neural network with "many" neurons and "many" layers
Smart: a neural network whose structure is based on the simulation algorithm

And compare their performance.

Vayel commented 6 years ago

Try other models (SVM, decision tree, etc.).

Vayel commented 6 years ago

The loss is stuck around 0.7 "whatever" the hyperparameters:

Loading datasets...
Training model...
1000: loss = 1.1729322919547558
2000: loss = 0.6851613330841064
3000: loss = 0.6740893872976303
4000: loss = 0.6692473928332329
5000: loss = 0.667147886633873
6000: loss = 0.6643219482600689
7000: loss = 0.6645239427685737
8000: loss = 0.6581559121608734
9000: loss = 0.6581234651505947
10000: loss = 0.6508811980485916
11000: loss = 0.6553445794284344
12000: loss = 0.6515623680651188
13000: loss = 0.6486354593634606
14000: loss = 0.6519274581670761
15000: loss = 0.6481850088536739
1000: loss = 0.890243825674057
2000: loss = 0.6483496547341346
3000: loss = 0.6462482803761959

It cannot understand the randomness. The network is deterministic but the simulation isn't.

Vayel / simudyne-back-end-test

Predict next breed of an agent #3