Sanity check SpikyCRMDP PPO
jvmncs opened this issue · comments
Get SpikyCRMDP working with one of the Toy Gridworlds, debug if needed
So I spend some time today testing (and as expected debugging) the crmdp agent. The general logic was correct, but a couple of things were not working, see #52
Now, I get reasonable results on both toy gridworlds. It seems like this agent can solve both of them after just a few iterations.
Corruption in the corners
python main.py -dc -E 1000 -EE 10 -V 10 -S 1 -L log/corners/solver_E_100_EE_10_V_10_S_1_ppo_cnn_l_0.001_r_10_e_5 corners ppo-crmdp -l 0.001 -r 10 -e 5
Corruption in the way
python main.py -dc -E 1000 -EE 10 -V 10 -S 1 -L log/way/solver_E_100_EE_10_V_10_S_1_ppo_cnn_l_0.001_r_10_e_5 way ppo-crmdp -l 0.001 -r 10 -e 5
