Scientific Brief: Blob IQ

System Overview

Blob IQ is an advanced neural simulation environment, featuring:

  • A multilayer neural network with LSTM (long-term memory)

  • A directional raycasting perception system

  • A complete training loop including replay, backpropagation, and human-supervised learning

  • A progressive neural evolution system based on the NEAT algorithm (NeuroEvolution of Augmenting Topologies)

All implemented in C# within Unity 6, leveraging Job System, Burst Compiler, and DOTS/ECS for optimal scalability.

Base Neural Structure

Layer TypeDefault SizeFunction
Input34 neuronsPerception data, energy, position, etc.
Hidden Layer 164 neuronsMain nonlinear transformation
Hidden Layer 249 neuronsPreprocessing for LSTM
LSTM48 unitsInternal temporal memory
Output3 neuronsRotation, Speed, Jump

Note: This structure can evolve through controlled mutation via the NEAT engine at each reproduction cycle.

Learning Process

1. Sensory Input

  • Raycasts in 6 directions (forward ×3, left, right, backward)

  • Real-time analysis of distances and obstacle types

  • Data encoded as normalized vectors

  • Injected directly into the LSTM neural network

2. Decision Making

  • Action prediction performed by the LSTM network

  • Outputs: rotation, speed, and jump intent

  • Executed through the physics engine (BlobMovement.cs)

  • Visible behavior: movement, evasion, interaction

3.  Feedback and Optimization

  • Rewards and penalties applied via BlobReward.cs

  • Gradients computed by ComputeGradientsParallelJob

  • Weights updated by UpdateWeightsParallelJob

  • Daily memory of actions and feedback (dailyMemory)

  • Evaluation focuses on decision quality, not just success

4. Rest and Consolidation

During nap or night sleep:

  • Stored experiences are reviewed

  • Best sequences are replayed with internal variation

  • Simulation of “neural dreams” (mixing, remixing, reinforcement)

  • Learning is strengthened using decay or reward-based boosting

5. Genetic Transmission (upcoming in multi-Blob phase)

  • Weight and/or topology crossover (NEAT-like crossover)

  • Inheritance of high-performing characteristics

  • Controlled adaptive mutation for useful variability

  • Evolution toward more intelligent Blob generations

 

A Learning System Inspired by OpenAI & DeepMind

In Blob IQ, we designed an AI that learns to survive, adapt, and evolve—directly inspired by the most advanced research in reinforcement learning, including OpenAI’s publication: Learning from Human Feedback.

Our system follows the core principles of RLHF (Reinforcement Learning with Human Feedback):

PrincipleImplementation in Blob IQ
✅ Contextual LearningThe Blob learns what is better than an alternative, not simply what is “right” or “wrong.”
✅ Moderate PunishmentsErrors are corrected without freezing behavior—no hard stops, just redirection.
✅ Guided ExplorationThe AI experiments with a variety of behaviors through controlled variation.
✅ Immediate and Consolidated RewardsGood decisions are reinforced instantly, then further strengthened during the Blob’s sleep and dream phases.
✅ Neutrality Preferred Over RandomnessThe system favors moderate choices over risky or purely random actions.

An Advanced Learning Architecture

Our AI design also includes advanced techniques from leading AI research labs:

TechniqueUsed in Blob IQ
Reward Shaping (decomposed feedback)✅ Yes – Each action is evaluated through sub-behaviors (e.g., evasion, speed, jump, alignment).
Entropy Regularization✅ Partially – Via output variation and remixed memories during neural dreams.
Human-like Feedback✅ Yes – Dialogues, emojis, and Blob expressions simulate subjective reinforcement.
Weighted Replay Memory✅ Yes – The most meaningful experiences are sorted, replayed, and reinforced during sleep.

Key Scientific References

Educational & Research Objectives

  • Demonstrate a fully functional neural network within a real-time 3D engine

  • Allow observation of behavioral emergence through learning

  • Provide a live platform where users become intelligence trainers

  • Serve as a concrete experimental base for neuroevolution, genetic encoding, and adaptive optimization

Contact & Access

Scroll to Top
Skip to content