Visrl

Visrl (pronounced "visceral") is a simple wrapper to analyse and visualise reinforcement learning agents' behaviour in the environment.

Reinforcement learning requires a lot of overhead code to inspect an agent's behaviour visually, typically through env.render(). Visrl allows users to easily intervene and switch between agent control and human control, and allows inserting a breakpoint in the game state to pause only at a relevant state of interest.

Features

• Set action hotkeys
• Human intervention: Take actions 1 step at a time
• Agent control: Return control to the agent
• Speed up/ slow down frame rate
• Visualise relevant values across history
• Breakpoint: Run until a condition involving values is fulfilled
• Playback: Show past frames and ations
• Record: Record a .mp4, .gif or download a .csv of the history.

Install

pip install visrl

Usage

import gym
from stable_baselines3 import DQN
from visrl import Visrl

env = gym.make('LunarLander-v2')
agent = DQN('MlpPolicy', env, verbose=1)
agent.learn(total_timesteps=int(2e5))

Visrl(env, agent).run()