Skip to content

Latest commit

 

History

History
33 lines (24 loc) · 2.64 KB

README.md

File metadata and controls

33 lines (24 loc) · 2.64 KB
Raw

trust_rl_VBA

VBA-style Q-learning and other models of modified Trust Game. A procedure to run programs in this collection can go as follows. (1) run trust_fit_group_vba. (2) After determining desired model, run trust_extract_PEs. (3) Run trust_write_PEs and trust_writevalues.

  1. trust_fit_group_vba.m: gets data from all subjects (.mat), runs a Qlearning (trust_Qlearning_ushifted) algorithm w/ preset parameters, saves F(it)-values in table.

2a. trust_Qlearning_ushifted.m: VBA fitting of Qlearning to a single subject trust data. Note that the u vector (which includes the subject's behavior +outcomes+experimental design) is shifted by one trial ahead, because enivronmental feedback on trial t-1 is what determines the value of the hidden states on trial t. Calls to f() and g() (obervation/evolution functions, respectively). Saves the figure and the results of the VBA inversion routine.

2b. trust_Qlearning_ushifted2.m: implemented for using with Social Value Model aka Fareri et al., 2015.

  1. Evolution functions: 3a. f_trust_Qlearn_counter.m: reinforcement reflects counterfactual rewards (with respect to the share action). 3b. f_trust_Qlearn1.m: reinforcement reflects counterfactual feedback from trustee (also with respect to the share actions). 3c. f_trust_Qlearn2.m: simpler evolution function where reinforcement for keep/share is considered separately. 3d. f_trust_Qlearn_counter_corrected.m: subject-counterfactual = actual rewards - would-be outcome from subject's alternative action
    3e. f_trust_Qlearn_counter_hybrid.m: subject-counterfactual counterfactuals are only calculated when participant's action = KEEP. 3e. f_trust_Qlearn_counter_trustee.m: trustee-counterfactual = actual rewards - would-be outcome from trustee's alternative action. 3f. f_trust_Qlearn_counter_hybrid_regret.m: regret = actual rewards - maximum outcome (1.5). 3g. f_trust_Qlearn_counter_hybrid_regret_purist: regret only when there is something to regret 3h. f_trust_SVM1: Social Value Model aka Fareri et al., 2015.

  2. Observation functions: 4a. g_trust_softmax_ED: observation function that includes the kappa parameter (bias for keep/share actions). 4b. g_trust_softmax_1: simpler observation function. 4c. g_trust_SVM1.m: Social Value Model aka Fareri et al., 2015 also softmax.

  3. trust_extract_PEs: gets the PEs from saved .mat files resulting from VBA inversion routines, saves in table.

  4. trust_write_PEs: writes regressors based on the table resulting from trust_extract_PEs.

  5. trust_writevalues: gets and writes value regressors.

  6. trust_modelComparisons: helper code to compares F(its) between models. Uses saved tables from trust_fit_group_vba.m.