RLSSM model config #805
Description
Introduce model configs for RLSSM models in HSSM to store information about the model parameters as well as other model/paradigm info such as choices, extra_fields, etc. In addition, other meta-data fields will need to be introduced to flag certain models as type RLSSM or getting the likelihood function name which would be implemented in some HSSM file such as rldm.py.
Current implementation
- See example model config below. We can reuse many fields from current SSM model configs but will have to add a few new fields.
- We can have a separate file that holds all the RLSSM model configs.
- We can potentially also add RL_process, SSM_process, RL_params and SSM_params attributes so that we have some metadata to construct likelihoods in a modular fashion. This would also help with downloading the LAN onnx for the corresponding SSM process. For example, RL_process can be 'Q-learning' and we define params that correspond to the Q-learning process. SSM_process can be 'ddm' and we can directly import ddm LAN.
- Since RLSSM models can be tricky for fit specially in terms of the priors, etc., we can also introduce a default priors attribute in the config. Or are the priors defined elsewhere in HSSM?
Example model config
"model_name": "custom_model_1"
"description": "<some text description>",
"type_rlssm": True,
"list_params": [param1, param2, ... paramN],
"bounds":{
"param1": (-inf, inf),
"param2": (0.0, 1.0),
...
"paramN": (1.0, 5.0),
},
"params_default": [1.2, 0.5, ... 2.4],
"data": ["rt", "response"],
"choices": [0, 1, 2],
"likelihood_function": "<function name from rldm.py>",
"extra_fields": ["ef_1", "ef_2", ... "ef_M"],