Possible bug: state visitation frequency

[magnusja]

Hey there,

I am not a 100% sure but I feel like there is something wrong with calculating the state visitation frequency (https://github.com/stormmax/irl-imitation/blob/master/deep_maxent_irl.py#L93).

You iterate over all the states and calculate the frequency for every timestep then.

for s in range(N_STATES):
    for t in range(T-1):
      if deterministic:
        mu[s, t+1] = sum([mu[pre_s, t]*P_a[pre_s, s, int(policy[pre_s])] for pre_s in range(N_STATES)])
      else:
mu[s, t+1] = sum([sum([mu[pre_s, t]*P_a[pre_s, s, a1]*policy[pre_s, a1] for a1 in range(N_ACTIONS)]) for pre_s in range(N_STATES)])

In my opinion the loops should be switched:

for t in range(T-1):
    for s in range(N_STATES):
      if deterministic:
        mu[s, t+1] = sum([mu[pre_s, t]*P_a[pre_s, s, int(policy[pre_s])] for pre_s in range(N_STATES)])
      else:
mu[s, t+1] = sum([sum([mu[pre_s, t]*P_a[pre_s, s, a1]*policy[pre_s, a1] for a1 in range(N_ACTIONS)]) for pre_s in range(N_STATES)])

Because the visitation frequency of timestep t+1 depends on all the state frequencies of timestamp t. This also reflects the formular from the original MaxEnt paper (Ziebart et al, 2008):

Unfortunately if I change the loop heads, the reward is not recovered correctly anymore. Do you have any hints on this?

[Zhousiyuhit]

Hello, I have encountered the same question as you. Have you solved it?

[magnusja]

Hello there,

please refer to my fork of this repository, which not only fixes that but also implements highly efficient methods for calculating the state visitation frequency, in tf but also vectorized using numpy. The code in this repository is completely unusable when you need more states than the 5 by 5 example grid ;D

The trick to fix the bug is essentially to take the average over timestamps. This is not mentioned anywhere except this video: https://youtu.be/d9DlQSJQAoI?t=973 (watch for a minute or so then Chelsea mentions that the calculation is missing an average).

See this note of mine as well: https://github.com/magnusja/irl-imitation/blob/master/deep_maxent_irl.py#L340-L348

Let me know if you have further questions.

[Zhousiyuhit]

Thanks very much~

在 2019年9月18日，15:11，Magnus [email protected] 写道：

Hello there,

please refer to my fork of this repository, which not only fixes that but also implements highly efficient methods for calculating the state visitation frequency, in tf but also vectorized with numpy. The code in this repository is completely unusable when you need more states than the 5 by 5 example grid ;D

The trick to fix the bug is essentially to take the average over timestamps. This is not mentioned anywhere except this video: https://youtu.be/d9DlQSJQAoI?t=973 https://youtu.be/d9DlQSJQAoI?t=973 (watch for a minute or so then Chelsea mentions that the calculation actually is missing an average).

See this note of mine as well: https://github.com/magnusja/irl-imitation/blob/master/deep_maxent_irl.py#L340-L348 https://github.com/magnusja/irl-imitation/blob/master/deep_maxent_irl.py#L340-L348 — You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/yrlu/irl-imitation/issues/1?email_source=notifications&email_token=AF3E3TV3COOJBSU4V2VZF6TQKHIBNA5CNFSM4EFFBU22YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD67BUPA#issuecomment-532552252, or mute the thread https://github.com/notifications/unsubscribe-auth/AF3E3TTE2JD5YWNLO6MYPRLQKHIBNANCNFSM4EFFBU2Q.

[Zhousiyuhit]

I modified the code based on tensorflow 2.0, and now there are no other problems.