pith. sign in

Verifiable Reinforcement Learning via Policy Extraction

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it
abstract

While deep reinforcement learning has successfully solved many challenging control tasks, its real-world applicability has been limited by the inability to ensure the safety of learned policies. We propose an approach to verifiable reinforcement learning by training decision tree policies, which can represent complex policies (since they are nonparametric), yet can be efficiently verified using existing techniques (since they are highly structured). The challenge is that decision tree policies are difficult to train. We propose VIPER, an algorithm that combines ideas from model compression and imitation learning to learn decision tree policies guided by a DNN policy (called the oracle) and its Q-function, and show that it substantially outperforms two baselines. We use VIPER to (i) learn a provably robust decision tree policy for a variant of Atari Pong with a symbolic state space, (ii) learn a decision tree policy for a toy game based on Pong that provably never loses, and (iii) learn a provably stable decision tree policy for cart-pole. In each case, the decision tree policy achieves performance equal to that of the original DNN policy.

fields

cs.LG 1

years

2019 1

verdicts

UNVERDICTED 1

representative citing papers

An Inductive Synthesis Framework for Verifiable Reinforcement Learning

cs.LG · 2019-07-16 · unverdicted · novelty 6.0

The paper introduces an inductive synthesis framework that generates verifiable deterministic program approximations of neural RL policies, preserving safety invariants via counterexample-guided search over state transition systems.

citing papers explorer

Showing 1 of 1 citing paper.

  • An Inductive Synthesis Framework for Verifiable Reinforcement Learning cs.LG · 2019-07-16 · unverdicted · none · ref 10 · internal anchor

    The paper introduces an inductive synthesis framework that generates verifiable deterministic program approximations of neural RL policies, preserving safety invariants via counterexample-guided search over state transition systems.