'type' must be changed to string ('int') instead of class type (int).
Because tensorforce/models/distribution_model.py checks for elif action['type'] == 'int' and as it is now, this test fails and self.distributions stays empty. This leads to a runtime error during a later point, e.g. when calculating log_probs for self.distributions for the gradients.
The tensorforce agent is currently not compatible with the v2 environment because of a small mistake:
https://github.com/MultiAgentLearning/playground/blob/4ba44f601157ab04aa0291bed7c2de4bde9e8f6b/pommerman/agents/tensorforce_agent.py#L27
'type' must be changed to string ('int') instead of class type (int).
Because tensorforce/models/distribution_model.py checks for
elif action['type'] == 'int'
and as it is now, this test fails andself.distributions
stays empty. This leads to a runtime error during a later point, e.g. when calculatinglog_probs
forself.distributions
for the gradients.