https://arxiv.org/abs/2007.04578
Readme was written by Dongjae Kim (kim10481@kaist.ac.kr)
Codes were written by Dongjae Kim (kim10481@kaist.ac.kr)
Two PFC-RL models are pre-trained as in the previous research (Lee et al., 2014). Since the models require 4 or 6 pre-trained parameters (unlike neural network's large number of weights), we just simply save it in csv formant for efficiency (regdata_*.csv files).
To generate simulation data of pfc-RL model1:
python main_gen_simul_task5.py --model-id=2 task-id=%d --sub-id=%d --GRW=%d --tree=%d
To generate simulation data of pfc-RL model2:
python main_gen_simul_task5.py --model-id=3 task-id=%d --sub-id=%d --GRW=%d --tree=%d
Codes were written by Minsu Abel Yang and Dongjae Kim (kim10481@kaist.ac.kr)
We have uploaded the trained weights of PM-metaRL (./meta-RL/PM/SUBFIT).
To generate simulation data of metaRL, simply run:
python main_gen_simul_task_meta_RL_PM.py --task-id=%d --sub-id=%d --GRW=%d --tree=%d
Codes were written by Jaehoon Shin (skalclrptsp@kaist.ac.kr)
We have uploaded the trained weights of PM-metaRL (./meta-RL/PM/SUBFIT).
To generate simulation data of metaRL, simply run:
python main_gen_simul_task_sbjtv_dqn.py --bhv_pseudo=0 --no_fitting=1 --fix_behavior=0 --task-id=%d --model-id=%d --sub-id=%d --GRW=%d --tree=%d --episodes=%d
Codes are written by Minsu Abel Yang and Dongjae Kim (kim10481@kaist.ac.kr)
We have uploaded the trained weights of PM-metaRL (./meta-RL/PM/SUBFIT).
To generate simulation data of metaRL, simply run:
python main_gen_simul_task_meta_RL_GM.py --task-id=%d --sub-id=%d --GRW=%d --tree=%d
Codes were written by Jaehoon Shin (skalclrptsp@kaist.ac.kr)
We have uploaded the trained weights of PM-metaRL (./meta-RL/PM/SUBFIT).
To generate simulation data of metaRL, simply run:
python main_gen_simul_task_sbjtv_dqn.py --bhv_pseudo=1 --no_fitting=1 --fix_behavior=0 --task-id=%d --model-id=%d --sub-id=%d --GRW=%d --tree=%d --episodes=%d
Description of arguments to specify tasks.
--sub-id: subject id. We used 82 subjects paramter sets so (0 to 81)
--model-id: Type of agent models. ex) 4: DDQN to maximize agents' rewards. 5: DDQN to copy the behavior of agents.
--bhv_pseudo: The sequence of goal conditions in the trainig. 0: Do not use pseuod sequence and use real subejcts sequence (model-id=5), 1: Do the pseudo sequence with equal numbers of each scenario (mode-id=4)
--no-fitting: Update DDQN or fix DDQN. 0: Update agents. 1: Freeze agents
--episodes: Number of episodes in training
--fix-behavior: The computational agents' sequences of visited states are exactly same as subjects' real data or not 0: Agents move as the agents select. 1: Agents move as the human subject's history.
--task-id : Type of tasks (1~10). ## the order is different from the order on the paper. please check arguments before running codes.
--GRW: the method to control state transition uncertainty using Gaussian random walk process or not. 0: switch-switch model, 1: fixed-drift, 2: drift-drift 3: fixed-(drif+switch) 4: (drift+switch)-(drift+switch)
--tree: Tree or ladder shape of the task 0: ladder, 1: tree