Safe Reinforcement Learning Algorithms

---

• HCOPE
• Safe Exploration
• Off Policy Evaluation
• Solving side effects

---

HCOPE (High-Confidence Off-Policy Evaluation.)

Python Implementation of HCOPE lower bound evaluation as given in the paper: Thomas, Philip S., Georgios Theocharous, and Mohammad Ghavamzadeh. "High-Confidence Off-Policy Evaluation." AAAI. 2015.

Requirements

• PyTorch
• Numpy
• Matplotlib
• scipy
• gym

Running Instructions

1. Modify the environment in the main function, choosing from OpenAI gym. (Currently the code works for discrete action spaces)
2. Run python hcope.py

Notes

• The file policies.py contains the policy used in the code. Modify the policy to suit your needs in this file.
• To reproduce the graph given in the original paper explaining the long tail problem of importance sampling, use the

visualize_IS_distribution()

method. Also, a graph of distribution of Importance sampling ratio is created which nicely explains the high variance of the simple IS estimator.

• All the values required for offpolicy estimation are initialized in the HCOPE class initialization.

• Currently the estimator policy is defined as a gaussian noise(mean,std_dev) added to the behavior policy for estimator policy initialization in the function setup_e_policy(). The example in paper uses policies differing by natural gradient. But, this works as well.

• To estimate c*, I use the BFGS method which does not require computing hessian or first order derivative.

• The hcope_estimator() method also implements a sanity check, by computing the discriminant of the quadratic in parameter delta(confidence). If it does not satisfy the basic constraints, the program prints the bound predicted is of zero confidence.

• The random variables are implemented using simple importance sampling. Per-decision importance sampling might lead to better bounds and is to be explored.

• A bilayer MLP policy is used for general problems.

• Results:
  Output format:

---

Safe exploration in continuous action spaces.

Paper: Safe Exploration in Continuous Action Spaces - Dalal et al.

Running Instructions

• Go inside safe_exploration folder
• First learn the safety function by collecting experiences
  python learn_safety_function.py
• Now using the learned safety function, add the path of these learned torch weights in the train_safe_explorer.py file. After that:
  python train_safe_explorer.py
  This enables agent to learn while following the safety constraints.

Results

• Safe exploration in a case where constraint is on crossing the right lane marker.

• Instability is observed in safe exploration using this method. Here constraint is activated going left through the center of the road.(0.3)

Explanation

• Linear Safety Signal Model

• Safety Layer via Analytical Optimization

• Action Correction

---

Importance Sampling

Implementation of:

• Simple Importance Sampling
• Per-Decision Importance Sampling
• Normalized Per-Decision Importance Sampling (NPDIS) Estimator
• Weighted Importance Sampling (WIS) Estimator
• Weighted Per-Decision Importance Sampling (WPDIS) Estimator
• Consistent Weighted Per-Decision Importance Sampling (CWPDIS) Estimator

Comparision of different importance sampling estimators:

Image is taken from phD thesis of P.Thomas:
Links: https://people.cs.umass.edu/~pthomas/papers/Thomas2015c.pdf

---

Side Effects

Penalizing side effects using relative reachability

Code - https://github.com/hari-sikchi/safeRL/tree/safe_recovery/side_effects

• Added a simple example for calculating side effects as given towards the end of paper

The relative reachability measure

Paper: Penalizing side effects using stepwise relative reachability - Krakovna et al.