Mava
Mava copied to clipboard
[BUG] get_agent_type assumes naming convention that is not consistent across all of our environments
Describe the bug
https://github.com/instadeepai/Mava/blob/develop/mava/specs.py#L68 assumes that agents from all environments are named with the convention type_identifier
and hence using split
to extract type. This is in some environments such as OpenSpiel/Tic Tac Toe since agents are named player_0 and player_1
To Reproduce
Steps to reproduce the behavior: 1.Run examples/tf/openspiel/tic_tac_toe/feedforward/decentralised/run_madqn.py 2.Log output of get_agents_types()
Expected behavior
Player_0 and Players_1 should be regarded as separate types and not a single agent type of name Player
Context (Environment)
- Latest Mava as of 24-02-22
Additional context
Possible Solution
Flag to specify environment type which will be used to set agent type