Home

🧭 IgnitionAI - Project Roadmap

This document outlines the phased development of IgnitionAI — a modular, browser-friendly framework for intelligent agent simulation and reinforcement learning.

---

✅ Phase 1 — Core Logic (MVP)

⚙️ Goal: Run agent-environment logic headlessly (no UI)

✅ Roadmap for "RL algo first" Phase A — @ignitionai/backend-tfjs only Implementing classic algorithms with TensorFlow.js

1. 🔁 Q-learning (tabular) – minimalist JS version without neural networks ✅ Implemented Q-Table agent with state/action lookup ✅ Added tests for basic functionality

2. 🧠 DQN – Deep Q-Network ✅ Implemented MLP simple input → hidden → output ✅ Added replay buffer with experience sampling ✅ Implemented target network with periodic updates ✅ Added epsilon-greedy exploration/exploitation ✅ Loss function based on TD error ✅ Unit tests with training validation

3. 🧘‍♂️ PPO – Policy Gradient ✅ Created initial PPO agent skeleton

• Implement Actor-Critic model
• Implement episode-based training
• Add policy and value loss functions

---

🚀 Phase 1.5 — Backend Infrastructure

🧰 Goal: Create robust, multi-environment backend support

✅ Created modular monorepo structure ✅ Implemented robust backend selection system ✅ Added support for all major TensorFlow.js backends:

• WebGPU (experimental)
• WebGL
• CPU
• WASM ✅ Added helper utilities for backend detection and info ✅ Comprehensive unit tests and error handling
• Add model serialization and loading

---

🌐 Phase 2 — R3F Visualisation

🎮 Goal: Make the agent & target visible in a 3D scene

• @ignitionai/r3f: add AgentMesh, TargetMesh, useAgent
• @ignitionai/demo-target-chasing: setup Vite + R3F scene
• Display step count and reward in the UI

---

🤖 Phase 3 — TFJS Backend (Training & Inference)

🧠 Goal: Train and run a model directly in the browser

✅ @ignitionai/backend-tfjs: built simple MLP model with configurable layers ✅ Implemented train() and predict() APIs via DQN agent

• Add model serialization with save() and load()
• Create streamlined Agent class interface

---

⚡ Phase 4 — ONNX Runtime Backend (Inference-only)

⚡ Goal: Run optimized pre-trained models in production

✅ Created initial package structure for ONNX backend

• Implement ONNX Runtime Web integration
• Add .onnx model loading and inference
• Create InferenceBackend wrapper

---

🧠 Phase 5 — Cognitive Agents (LLM & Planning)

📚 Goal: Enable agents with memory, reasoning and goals

• @ignitionai/backend-langchain: LLM-powered agent
• @ignitionai/backend-vercelai: edge-deployed AI actions
• Add simple text environment or RAG-driven simulation

---

🏁 Stretch Goals

• Physics-based environment (BallBalancer, CartPole3D)
• Multi-agent mode
• Export agent replay logs
• Web UI training dashboard
• OpenHub-like demo launcher

---

Built with ❤️ by Salim (@IgnitionAI)