Evaluating Policies

Intro

We can evaluate policies to examine the progress of the learning algorithms.

Method

1. to evaluate them automatically, use scripts/poli_eval/poli_val.py
2. ​you will have to point it to the directory containing the intermediate files using poli_files_dir
3. and you'll have to provide it with an arg file that specifies the argumenets
4. ​the default is args/opt_int_poli_eval.txt
5. but you should probably use a separate one for the hopper
6. ​you'll also specify the output file in the arg file
7. the output of the script will be a text file where each line contains the distance traveled for each individual expisode for a given policy
8. ​the default uses 100 episodes, so there should be 100 numbers per line
9. ​to generate the plots
10. ​use scripts/poli_eval/plot_eval.m
11. ​you just have to add the path to the eval output
12. ​and it should generate the training curve automatically