Initially, I fitted a minimal RW model using the fmincon function, and the code for these fits referenced some other repositories.
Then I found this toolbox and re-fitted the same RW model using a response model of type softmax_binary. But I found a big difference in the results. Mainly the beta parameter seems to be limited to the 0-1 range, which can be set to 0-10 in fmincon, and the learning rates are different. I can't figure out where these differences come from, and checking the implementation of this huge toolbox is too difficult for me. I'm guessing it stems from differences in the optimization function. Or is it because beta is explicitly limited to 0-1 in the toolbox?
Initially, I fitted a minimal RW model using the fmincon function, and the code for these fits referenced some other repositories. Then I found this toolbox and re-fitted the same RW model using a response model of type softmax_binary. But I found a big difference in the results. Mainly the beta parameter seems to be limited to the 0-1 range, which can be set to 0-10 in fmincon, and the learning rates are different. I can't figure out where these differences come from, and checking the implementation of this huge toolbox is too difficult for me. I'm guessing it stems from differences in the optimization function. Or is it because beta is explicitly limited to 0-1 in the toolbox?
Thanks to anyone who can answer this question.