Follow-up to: "I Implemented Deep Q-Networks in Pure PowerShell 5.1"
When I wrote the original article, VBAF was primarily a commercial automation engine. A lot has changed.
What changed:
VBAF has been repositioned as a purely educational framework. The commercial layer is gone. The focus is now on making AI and reinforcement learning concepts accessible to anyone with a Windows PC.
What is new in v5.0:
Full educational comments across 16 core files. Not just "what the code does" but WHY -- the Bellman equation with every symbol explained, the PPO clip trick with intuition, GAE lambda tradeoff, n-step returns with bootstrapping.
Three complete RL algorithms implemented and benchmarkable head-to-head:
$dqn = (Invoke-DQNTraining -Episodes 100 -FastMode)[-1]
$ppo = (Invoke-PPOTraining -Episodes 100 -FastMode)[-1]
$a3c = (Invoke-A3CTraining -Episodes 100 -FastMode)[-1]
$env = New-VBAFEnvironment -Name "CartPole"
Invoke-VBAFBenchmark -Agent $dqn -Environment $env -Episodes 20 -Label "DQN"
Invoke-VBAFBenchmark -Agent $ppo -Environment $env -Episodes 20 -Label "PPO"
Invoke-VBAFBenchmark -Agent $a3c -Environment $env -Episodes 20 -Label "A3C"
Multi-agent market simulation with 4 competing companies. Emergent behaviours -- price wars, tacit collusion, innovation races -- appear without being programmed. The Herfindahl index is computed at the end to measure market concentration.
Teaching materials in docs/teaching/ -- 4-week course outline, lab exercises, exam questions.
Why PowerShell?
Because the code is readable. You can open any .ps1 file and see exactly what the algorithm is doing. No black boxes. No library abstractions. The code IS the textbook.
Install-Module VBAF | GitHub: https://github.com/JupyterPS/VBAF
"The best way to understand AI is to build it yourself -- line by line."
Top comments (0)