Skip to content

Experiments

Kevin edited this page Jan 16, 2019 · 54 revisions

Here we can plan experiments and gather their results. The tasks that we consider (all have the dimension 30x30 pixel):

  • Evasion (Obstacle chance = 0.05) (max Score 5000)
  • Race (Obstacle chance = 0.05) (max Score 5000)
  • Tunnel (max Score 5000)
  • EvasionWalls (Obstacle chance = 0.02) (max Score 5000)
  • VisualObstaclePathing
  • VisualSimplePathing
ID Environments Representation learner Policy learner Episodes Average max score Run by Running Done Time
1 Evasion CerberusPixel (n_hidden=30) DoubleDQN (eps_decay=2000) 1000 356 Kevin [x] [x] 5min
2 Evasion Flatten DoubleDQN (eps_decay=2000) 10000 448 Kevin [x] [x] 123min
3 Evasion CerberusPixel (n_hidden=30) DoubleDQN (eps_decay=2000) 10000 362 Kevin [x] [x] 68min
4 Evasion Flatten DoubleDQN (eps_decay=2000) 5000 (-1000 ckpt) 482 Alessandro [x] [x] 12min
5 Evasion Flatten DoubleDQN (eps_decay=2000) 10000 (-5000 ckpt) 1160 Alessandro [x] [x] 19min
6 Evasion Flatten DoubleDQN (eps_decay=2000) 15000 (-10000 ckpt) 502 Alessandro [x] [x] 28.9 min
7 Evasion VariationalAutoencoderPixel (n_hidden=30) DoubleDQN (eps_decay=2000) 100000 482 Alessandro [x] [x] 8hrs
8 Tunnel Flatten DoubleDQN (eps_decay=2000) 10000 456 Kevin [x] [x] 107min
9 EvasionWalls CerberusPixel (n_hidden=30) DoubleDQN (eps_decay=10000, memory_delay=10000, FullBackprop) 10000 378 Kevin [x] [x] 150min
10 Tunnel CerberusPixel (n_hidden=30) DoubleDQN (eps_decay=10000, FullBackprop) 10000 163 Kevin [x] [x] 37min
11 Evasion Flatten DoubleDQN (eps_decay=2000) 40000 1231 Alessandro [x] [x] 3.25hrs
12 Tunnel Flatten DoubleDQN (eps_decay=1500) 20000 290 Alessandro [x] [x] 40min
13 EvasionWalls Flatten DoubleDQN (eps_decay=20000, memory_delay=100000) 30000 442 Alessandro [x] [x] 110min
14 EvasionWalls Flatten DoubleDQN (eps_decay=15000) 100000 583 Alessandro [x] [x] 8.3hrs
15 Race Flatten DDQN (epsdecay=2000, lr=5.e-4) 10000 541.5 Tonio [o] [x] 30.91min
16 Tunnel Flatten DeepQNetwork 50000 581 Tonio [x] [x] 105
17 EvasionWalls Flatten DoubleDeepQNetwork 50000 779.0 Danni [x] [x] 464.34
18 Tunnel ConvolutionalPixel DoubleDeepQNetwork 100000 196.8 Alessandro [x] [x] 269min
19 Tunnel JanusPixel (n_hidden=30) DoubleDQN (eps_decay=10000, FullBackprop) 50000 Kevin [x] []

Clone this wiki locally