andreicnica/hrl_attention

The Paradox of Choice: Using Attention in Hierarchical Reinforcement Learning

Using Attention in HRL

Framework for training options with different attention mechanism and using them to solve downstream tasks.

Requirements

GPU required

conda env create -f conda_env.yml

After the instalation ends you can activate your environment and install remaining dependencies. (e.g. sub-module gym_minigrid which is a modified version of MiniGrid )

conda activate affenv
cd gym-minigrid
pip install -e .
cd ../
pip install -e .

Instructions

In order to train options and IC_net follow these steps:

1. Configure desired environment - number of task and objects per task in file config/op_ic_net.yaml. E.g:
  env_args:
    task_size: 3
    num_tasks: 4

2. Configure desired type of attention (between "affordance", "interest", "nan") - in file config/op_ic_net.yaml. E.g. 
main:
  attention: "affordance" 

3. Train by running command
liftoff train_main.py configs/op_ic_net.yaml

Once a pre-trained option checkpoint exists a HRL agent can be trained to solve the downstream task (for the same environment the options were trained on). Follow these steps in order to train an HRL-Agent with different types of attentions:

1. Configure checkpoint (experiment config file and options_model_id) for pre-trained Options and IC_net - in file configs/hrl-agent.yaml. E.g: 

main:
  options_model_cfg: "results/op_aff_4x3/0000_multiobj/0/cfg.yaml"
  options_model_id: -1  # Last checkpoint will be used

2. Configure type of attention for training the HRL-agent (between "affordance", "interest", "nan") - in file configs/hrl-agent.yaml. E.g:
main:
  modulate_policy: affordance

3. Train HRL-agent by running command
liftoff train_mtop_ppo.py configs/hrl-agent.yaml

Both training scrips produce results in the results folder, where all the outputs are going to be stored including train/eval logs, checkpoints. Live plotting is integrated using services from Wandb (plotting has to be enabled in the config file main:plot and user logged in Wandb or user login api key in the file .wandb_key).

The console output is also available in a form:

• Option Pre-training e.g.:

U 11 | F 022528 | FPS 0024 | D 402 | rR:u, 0.03 | F:u, 41.77 | tL:u 0.00 | tPL:u 6.47 | tNL:u 0.00 | t 52 | aff_loss 0.0570 | aff 2.8628 | NOaff 0.0159 | ic 0.0312 | cnt_ic 1.0000 | oe 2.4464 | oic0 0.0000 | oic1 0.0000 | oic2 0.0000 | oic3 0.0000 | oPic0 0.0000 | oPic1 0.0000 | oPic2 0.0000 | oPic3 0.0000 | icB 0.0208 | PicB 0.1429 | icND 0.0192

Some of the training entries decodes as

F - number of frames (steps in the env)
tL - termination loss
aff_loss - IC_net loss
cnt_ic - Intent completion per training batch 
oicN - Intent completion fraction for each option N out of Total option N sampled
oPicN - Intent completion fraction for each option N out of affordable ones
PicB - Intent completion average over all options out of affordable ones

• HRL-agent training

U 1 | F 4555192.0 | FPS 21767 | D 209 | rR:u, 0.00 | F:u, 8.11 | e:u, 2.48 | v:u 0.00 | pL:u 0.01 | vL:u 0.00 | g:u 0.01 | TrR:u, 0.00

Some of the training entries decodes as

F - number of frames (steps in the env offseted by the number of pre-training steps)
rR - Accumulated episode reward average
TrR - Average episode success rate

Framework structure

The code is organised as follows:

• agents/ - implementation of agents (e.g. training options and IC_net multistep_affordance.py; hrl-agent PPO ppo_smdp.py )
• configs/ - config files for training agents
• gym-minigrid/ - sub-module - Minigrid envs
• models/ - Neural network modules (e.g options with IC_net aff_multistep.py and CNN backbone extractor_cnn_v2.py)
• utils/ - Scripts for e.g.: running envs in parallel, preprocessing observations, gym wrappers, data structures, logging modules
• train_main.py - Train Options with IC_net
• train_mtop_ppo.py - Train HRL-agent

Acknowledgements

We used PyTorch as a machine learning framework.

We used liftoff for experiment management.

We used wandb for plotting.

We used PPO adapted for training our agents.

We used MiniGrid to create our environment.