Skip to content

Latest commit

 

History

History
87 lines (56 loc) · 5.03 KB

README.md

File metadata and controls

87 lines (56 loc) · 5.03 KB
Raw

RL, Policy Gradients, Imitation Learning and Bootstrapping

This Project aimed to train models for MuJoCo Environments - Hopper, Half-Cheetah and Ant - in increasing order of complexity and number of joints.

More details about the environments and control parameters can be found at here!

The inituitive difficulty of these models is the requirement to fit a continuous action space. I take three approaches to the problem, an Imitation Learning Agent, a Policy Gradient Based Approach (both vanilla Actor Critic and Soft Actor Critic), and lastly building a better SAC model by bootstrapping it via IL.

hopper

Setup

Dependencies are listed in the environment_lin.yaml file. I suggest creating a conda environment and installing these using pip install -e .

If unsuccesful, install by executing :

  1. conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
  2. pip install -r requirements.txt
  3. pip install -e .

Running Instructions

  1. The Agents are housed in agents/
  2. The script for testing and training the model are in scripts/
  3. Best models are automatically saved to best_models/
  4. utils/ houses functionality for buffers and logging.

To run the training loops, execute python scripts/train_agent.py --env_name <ENV> --exp_name <ALGO> [optional tags] where,

  • <ENV> is one of Hopper-v4, HalfCheetah-v4 or Ant-v4
  • <ALGO> is one of imitation, RL or imitation-RL
  • [optional tags] like no_gpu and log frequencies can be used to tune runtime environment.

Imitation Learning Agent

Key Insights and Implementational details of Imitation Agent:

  • NN Model : Simple MLP , Learning Rate = 1e-3,
  • Beta (DAGGER Constant) : Decreases as itr_num increases ~ (1/1+(itr_num))
  • With Probability Beta, train on Expert Policy, with 1-Beta, train on your action
  • After every iteration, if model performs better than previous best, checkpoint it!
  • For every training iteration, randomly sample 10 trajectories from the buffer and retrain on them.
  • Best Performance : Hopper <2353>, HalfCheetah <2966>, Ant <1396>

Reinforcement Learning Agent:

  • Continuous Action space is handled by assuming that all dimensions of actions are independent of each other and using a (not too unrealistic) assumption that the actions can be assumed to be coming from a Normal Distribution

  • Thus, the model outputs 2*action_dim number of values, half being the means and the other half being variances

  • When querying for the next action, a sample from this distribution is returned

  • Vanilla Reinforce is just too unstable

  • With Baseline, doesn't seem to train on anything at all

  • With Actor Critic, learns and (unstably) stays for small patches on a small reward (~1000).

  • Does not explore enough to start hopping though

  • Every training iteration, a random sample of 10 trajectories from the replay buffer is re-trained to combat forgetfulness

  • I also incentivise higher traj lengths : by injecting reward proportional to remaining length of trajectory for that state

  • The rationale behing this is that for the given (s), the model could run for (T-t) more timesteps, and if T-t is higher, the mode is able to run more

  • Gradients that are added to the current policy are weighed by alpha : sqrt(1/(1 + envsteps_so_far/1000))

  • This decreases as iterations increase, inspiration from simulated-annealing.

  • Implementation of Soft Actor Critic : Theoretic results from SAC suggest taking minimum of multiple critics to combat over estimation, soft-transition of parameters (tau), representation (reward_now * gamma * Q(s,a) - V(s)) of Advantage, and maximizing "entropy"

  • Implementation With help from SAC

  • Soft-Updates of parameters for stability, multi-critics to combat over-estimation

Imitation-Seeded RL

Key Insights and Implementational Details from ImitationSeededRL:

  • In the first Iteration of training, I train the Imitation attribute ("the imitator") to fit the expert policy well

  • Once trained, I fit the actor model to actions suggest by the imitaton on sampled trajectories, and the critic to the "rewards-to-go" obtained from these trajectories on a per-state basis.

  • The ImitationSeededRL version of the Soft Actor Critic is purely my contribution.

  • On subsequent iterations, with probability p, perform some training iterations on the imitator and "recalibrate" the actor and critic to the imitator

  • With probability 1-p, let the RL component of the model explore

  • As the number of iterations increase, decrease p (explore more).

Plots and Logs

  • Greedily save the model when a test set performs better than the previous best performance

  • Analysis of performance has has been meticulously performed on the trainings. Some Plots of performance vs training have been attached in here