Skip to main content

No Representation, No Trust: Connecting Representation, Collapse, and Trust Issues in PPO

Created on October 29|Last edited on October 29

Figure 1



20M40M60M80Mcounters/env_steps102030405060
20M40M60M80Mcounters/env_steps100200300
20M40M60M80Mcounters/env_steps1100
20M40M60M80Mcounters/env_steps20004000600080001000012000
ALE/Phoenix
15
ALE/NameThisGame
15
MuJoCo Hopper
15
ALE
15
MuJoCo
15




Figure 1


ALE/Phoenix
30
ALE/NameThisGame
30
ALE
15
Hopper
30
MuJoCo
0




Figure 2



ALE/Phoenix
15
ALE/NameThisGame
15
MuJoCo Hopper
15
ALE
15
MuJoCo
15



Figure 3


NameThisGame
3
NameThisGame plasticity
3



Hopper 10
3
Hopper plasticity
3
HalfCheetah 10
1
HalfCheetah plasticity
1




Interventions


ALE
5
regularize features
3
share trunk
3
reset optim
3