Skip to content

Reinforcement learning

Behavioral cloning in Atari Pong

Imitation learning with a 3-layer CNN

Timeline
Sep 2024 to Oct 2024
Affiliation
Worcester Polytechnic Institute
Test accuracy
87.55%
Action latency
<500 ms

What it is

A supervised imitation-learning pipeline that clones an expert Pong policy from Arcade Learning Environment demonstrations with a three-layer convolutional network.

Problem

Can a small CNN reproduce expert paddle control from demonstrations alone without reward signals.

Approach

Trained on expert frames for 100 epochs, validated through loss trajectories and gameplay video logs, and measured action latency.

Highlights

  1. 87.55% test accuracy with minimal overfitting; the cloned agent tracks the paddle with sub-500 ms action latency.