Evolutionary Strategy
In this repo I tried using Evolutionary Stratgies to train a neural network to play Super Mario Bros. for the NES. The generall approach of this type of machine learning can be found here.
Trainingssession on SuperMarioBros-NES
Level 1-1 ✔/✔
| Training 28 uses Agent_old (no sprinting) | Training 42 uses the (new) Agent |
|---|---|
 |
 |
Level 1-2 ❌/✔
| Training 29 uses Agent_old (no sprinting) | Training 43 uses the (new) Agent |
|---|---|
 |
 |
Level 1-3 ❌/✔
| Training 41 uses the (new) Agent | Training 44 uses the (new) Agent |
|---|---|
 |
 |
Level 1-4 ✔
| Training 45 uses the (new) Agent |
|---|
 |
Level 1-1 and Level 1-2 in one Training ✔
| Training 48 uses the (new) Agent | From the same Trainingssession |
|---|---|
 |
 |