variable-td3

Traditionally, we learn a policy, and action is determined for every time-step. However, in many cases, it is also viable to simply repeat an action for multiple time-steps rather than determining a new action every time. This repeat factor is usually manually tuned and is kept constant. We hypothesize that keeping it constant may not be an ideal-policy as there could be scenarios in an environment where we need fine-step control as well as there could be scenarios where a larger-step control is feasible. For example, if we think of Lunar-Lander, we may need fine-step control as we are closer to the ground and attempting to land as compared to moments when we are high up in the space and large-repeat action may be feasible.

In this work, we learn a policy that learns an action as well as the time-step for which this action should be repeated. This gives the policy the ability to have large as well as fine-step control. We also hypothesize that learning to repeat an action may also lead to better sample efficiency. Our work utilizes "td3" as a core learning algorithm and updates it to have a q-value for each action-repeat, thereby, we call it "variable-td3".

Slide

Status:

• Work has been halted.
• Code provided as it is and no major updates expected.
• This work is similar to the paper "Learning to repeat: fine grained action repetition for deep reinforcement learning"

Installation

1. Install conda

2. For classic tasks, do following:

   conda env create -f  env.yml # creates env with name "vtd3"

3. Optional: For gym mujoco env. do following:

   • Requires mjpro 150 and mujoco license
   • conda activate vtd3 # activates conda-env. in (2)
   • pip install 'gym[mujoco]'

4. Optional: For dm_control, create a separate conda env. having mujoco 2.0 :

   • Requires mujoco200 and mujoco license
   • conda env create -f env_mj2.yml # creates env with name "vtd3_mj2"

Having trouble during installation ?, please refer here

Usage

• $ conda activate <env_name>

• Train: $ python main.py --case classic_control --env Pendulum-v0 --opr train

• Test: $ python main.py --case classic_control --env Pendulum-v0 --opr test

  Required Arguments	Description
  --case {classic_control,box2d,mujoco,dm_control}	It's used for switching between different domains(and configs)
  --env	Name of the environment Environments corresponding to ease case: classic_control : {Pendulum-v0, MountainCarContinuous-v0} box2d : {LunarLanderContinuous-v2, BipedalWalker-v3, BipedalWalkerHardcore-v3} mujoco: (refer here) dm_control: (refer here)
  --opr {train,test}	select the operation to be performed

• Visualize Results: tensorboard --logdir=./results

• Summarize plots in Plotly:

  $ cd scripts
  $ python summary_graphs.py --logdir=../results/classic_control --opr extract_summary 
  $ python summary_graphs.py --logdir=../results/classic_control --opr plot

Installation Troubleshooting

Windows :

• • Error: error: command 'swig.exe' failed ...
  • Fix: install swigy from here and add it in your path using this reference.
• • Error: error: Microsoft Visual C++ 14.0 is required. ...
  • Fix: by installing build tools from here.