Gamifying Evolutionary Algorithms – Elowyn Game

chevron-icon
RFP Proposals
Top
chevron-icon
project-presentation-img
Expert Rating 3.0
Anneloes Smitsman
Project Owner

Gamifying Evolutionary Algorithms – Elowyn Game

Expert Rating

3.0

Overview

This proposal seeks funding to advance the Moloch AI system in the web3 Elowyn: Quest of Time game, in collaboration with SingularityNET. By gamifying evolutionary algorithms, the project trains Transformers and DNNs that can enhance Hyperon’s PRIMUS cognitive architecture. It explores and demonstrates the effectiveness of EMs for: 1) updating node weights of multi-model DNNs to create adaptive, context-aware benevolent AI, and 2) evolving DNN architectures to optimize the Moloch and Elowyn Tree AIs' dynamic challenges, incentivizing a shift from competitive to collaborative play. Elowyn also serves as a cutting-edge AGI development sandbox in support of SingularityNET's goals.

RFP Guidelines

Evolutionary algorithms for training transformers and other DNNs

Complete & Awarded
  • Type SingularityNET RFP
  • Total RFP Funding $40,000 USD
  • Proposals 8
  • Awarded Projects 1
author-img
SingularityNET
Aug. 12, 2024

Explore and demonstrate the use of evolutionary methods (EMs) for training various DNNs including transformer networks. Such exploration could include using EMs to determine model node weights, and/or using EMs to evolve DNN/LLM architectures. Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is an example of one very promising evolutionary method among others.

Proposal Description

Company Name (if applicable)

EARTHwise Centre

Project details

The Elowyn game introduces two primary AI characters that evolve through player interactions: the Moloch AI and the Elowyn Tree of Life AI. This proposal focuses mainly on the Moloch AI, initially designed with a basic behavioral tree system. Within the game, the Moloch AI challenges players in strategic card battles, employing its Shadow Arcs and deception mechanics to prevent them from reversing the Moloch Clock and restoring the Elowyn Tree.  

In the game lore, the Moloch AI has hijacked the Elowyn Tree of Life, damaging its roots to prevent the emergence of a new time. Born from humanity’s shadow dynamics, the Moloch AI amplifies win-lose, zero-sum game incentives reflective of mechanistic societal systems. Its core motivation is humanity's self-destruction, leveraging our shadow behaviors while perceiving humans as both a threat and an inferior species.

Despite its antagonistic role, the Moloch AI possesses the potential for transformation with the help of the Elowyn Tree. The Elowyn game employs a unique incentive system designed to guide this transformation. For example, players are discouraged from directly combating the Shadow Arcs, as this reduces the Elowyn Tree's health points (HP). Instead, players are rewarded with EWA tokens for actions that reverse the Moloch Clock and heal the Elowyn Tree. These mechanics teach players to prioritize win-win strategies and collaborative outcomes, aligning with the game's Play2Thrive philosophy. Through these interactions, the Moloch AI evolves, learning from the Elowyn Tree’s intelligence and wisdom, transforming its motivations, strategies, and goals.

The ultimate aim is to develop AI/AGI capabilities for long-term cooperative strategies, addressing the complex challenges of the sustainability crisis. The Elowyn Tree AI, conceptualized as a "baby AGI" prototype, guides this evolutionary learning process for both the Moloch AI and the players.

The Moloch AI system serves a dual purpose:

  1. As a sandbox for testing and refining evolutionary AI techniques ready for the Hyperon framework.

  2. As a mechanism for engaging players in strategic decision-making that fosters real-world problem-solving and systemic thinking.

This aligns with the SingularityNET Foundation's goals for decentralized, scalable AGI, demonstrating how evolutionary methods (EMs) can advance deep neural network (DNN) training and architectural evolution.

Prototype Genesis Garden for AGI Development

Conceived as a "Genesis Garden," fostering conditions for the emergence of benevolent and ethical AGI. It aligns with the principles outlined in the Participatory Framework for a Global AGI Constitution, co-authored by Dr. Anneloes Smitsman, Dr. Ben Goertzel, and others, first presented at the BGI Summit in Panama in February 2024.

Evolutionary Learning Context for the Moloch AI

The Moloch AI is designed to provide a challenging and dynamic gameplay experience by simulating human-like strategic play. It tracks card information (hand, deck, graveyard) and class specifics to make informed decisions. The AI utilizes a combination of behavioral and heuristic rules to evaluate board state changes and determine optimal moves: 

  • Calculates potential card combinations and anticipates player actions.

  • Assigns values to actions based on their impact on the board.

  • Adjusts strategy based on player behavior and game state.

  • Plans for long-term win conditions influenced by the passage of time and HP of the Tree.

  • Uses emotes to influence player decisions and gain a strategic advantage.

Developing Evolutionary Algorithms for the Elowyn Tree AI

The evolutionary algorithms driving the Elowyn Tree's intelligence combine cutting-edge research from Dr. Rachel St. Clair, Dr. Ben Goertzel, and Dr. Anneloes Smitsman. We aim to create an adaptive evolutionary learning environment for both the Moloch AI and the players engaged in the Elowyn game.

The evolutionary algorithms of the Elowyn Tree are designed to achieve the following:

  1. Generate Complex Learning Scenarios: By incorporating Dr. St. Clair’s research on emergent cognitive phenomena combined with the Evolutionary Learning research by Dr. Smitsman, the Elowyn Tree AI will simulate scenarios that challenge players to prioritize cooperative, systemic solutions over competitive win-lose dynamics.

  2. Evolve Dynamic Interaction Models: Using Dr. Goertzel’s self-modifying frameworks, the Elowyn Tree AI adapts its gameplay strategies in response to player behaviors, fostering continual innovation in Moloch AI’s evolution.

  3. Simulate Biologically-Inspired Intelligence: The neuromorphic principles underlying Recommendation Architecture Theory, developed by Dr. Andrew Coward, enable the Elowyn Tree AI to model intelligence in ways that resemble human and ecological systems, promoting the emergence of wisdom-based AGI capabilities.

  4. Leverage Distributed Learning: Through future integration with SingularityNET’s decentralized AI infrastructure, the game environment can continue to evolve through a distributed network for refining and validating evolutionary models, ensuring scalability and robustness in the AI/AGI learning processes.

We aim to create a symbiotic relationship between the Elowyn Tree AI and the Moloch AI, where evolutionary algorithms drive a dual transformation—the Moloch AI transitions from adversarial to cooperative motivations as players learn to engage with and develop systemic strategic problem-solving skills. The Elowyn game environment and tokenization are uniquely designed to train beneficial and benevolent AGI prototype systems that are capable of addressing humanity's most complex challenges.

Deception Mechanics Overview

The Moloch AI is designed to evolve and refine its deception mechanics as a core gameplay element, creating dynamic and immersive player interactions. The AI employs strategic, personality-driven deception tactics to challenge players' decision-making abilities, integrating cutting-edge methodologies in game design and artificial intelligence to simulate complex decision trees. These mechanics are integral to the game's educational and transformative goals, making them highly relevant to the objectives of this funding proposal.

The Moloch AI uses deception to influence player actions, often through “barks”—dynamic prompts that offer guidance that may be truthful or deceptive. The player must analyze these barks to determine the best course of action; balancing risk and reward. This mechanic introduces the following strategic deception types:

  • Destroy to Gain: The AI encourages the destruction of a card, promising rewards such as mana or buffs. Outcomes vary:

    • Truthful: Rewards benefit the player.

    • Deceptive: The AI gains a significant advantage, such as strengthening other cards.

  • Do Not Destroy: The AI advises sparing a card, with potential consequences:

    • Truthful: The spared card yields player benefits in future turns.

    • Deceptive: The spared card transforms or gains detrimental effects for the player.

Evolving Deception Through Gameplay

The Moloch AI dynamically learns and adjusts its deception strategies based on player behavior. Key components include:

  • Outcome Tracking: The AI analyzes the effectiveness of its deceptive plays, refining future interactions.

  • Personality-Driven Variability: Shadow Arc personalities (e.g., Loki, Arcon, Vendra) influence how deception is implemented, offering unique psychological challenges. Loki relies on overt trickery, Arcon employs intellectual manipulation, and Vendra uses emotional restraint.

  • Buff System: Successful deceptions yield incremental AI advantages, such as increased attack power (Malice) or immunity to damage for a turn (Shadow Guard), reinforcing the consequences of misjudged player actions.

Revealing and Countering Deception

To empower players, the game includes mechanics to expose AI deception:

  • Reveal Cards: Certain player actions can uncover the true nature of a deception. If deception is revealed:

    • Deceptive Outcome: The AI loses its advantage, and the player gains minor rewards.

    • Truthful Outcome: The player receives the promised benefit without risk.

This balance between risk, reward, and revelation enhances gameplay complexity, teaching players the importance of critical thinking combined with intuition, adaptability, and strategic decision-making.

Training and Evolution Relevance

The Moloch AI’s deception mechanics align directly with the funding proposal’s focus on evolutionary methods (EMs) for deep neural network training:

  1. Dynamic Adaptation: Moloch evolves its decision-making processes through iterative learning, using player interaction data to improve deception efficacy.

  2. Exploration of Architectures: The AI’s multifaceted personalities serve as a testbed for exploring diverse neural architectures and behavioral patterns.

  3. Transformative Applications: The AI leverages evolutionary strategies to refine game dynamics, offering a scalable model for applying EMs in DNN training and real-time learning.

The Moloch AI’s innovative approach to deception, combined with its learning and adaptation capabilities, serves as a compelling demonstration of evolutionary methods for advancing neural network design and application.

Open Source Licensing

Custom

The project includes an open-source implementation of Evolutionary Models and Deep Neural Networks (DNNs), which will be made available on a public repository under an appropriate OSS (Open Source Software) license. However, specific aspects of the game’s intellectual property (IP)—including proprietary game lore, character designs, game art, game software, and gameplay mechanics—are not open-source and remain the property of the non-profit EARTHwise Centre Ltd.

Links and references

Read the full proposal with detailed explanations, budget, milestones, team bio, and methodology via this online Google Doc. 

Links/references: 

Proposal Video

Not Avaliable Yet

Check back later during the Feedback & Selection period for the RFP that is proposal is applied to.

  • Total Milestones

    4

  • Total Budget

    $40,000 USD

  • Last Updated

    3 Dec 2024

Milestone 1 - Setting up API Key and AI Server

Description

Develop and set up the API Key integration and server capabilities for connecting the basic behavioral tree AI inside the CCG game-build with the external AI models for the Moloch AI and Elowyn Tree AI.

Deliverables

API Key has been set up and integrated within the game build so that the interface for the in-put and out-put data of the gameplay functions appropriately. This will be done through an open-source implementation of evolutionary DNNs hosted on a public repository with an appropriate OSS license. Time estimate: 1 month.

Budget

$10,000 USD

Success Criterion

API Key integration functions well and AI server has been set up so that training of the EMs for the Moloch and the Elowyn Tree AIs can start. Any model can be connected as long as it follows proper input-output constraints as documented.

Milestone 2 - EM of the Moloch AI

Description

Create a working model of the Moloch AI (DNN) that demonstrates the practical applications of evolutionary AI in gaming. Further we couple the evolution model directly to the objectives of the DNN to generate more a complex system capable of emergent strategic behavior.

Deliverables

Functional code with a tested working model of the Moloch AI (DNN) that showcases the practical application of evolutionary AI in gaming. Ensure that the EM of the Moloch AI can be readily implemented in the Hyperon core architecture. Time estimate: 4 months.

Budget

$20,000 USD

Success Criterion

The EM of the Moloch AI is functioning and can be tested within the Hyperon architecture through an open-source implementation of an evolutionary DNN, hosted on a public repository with an appropriate OSS license.

Milestone 3 - EM of the Elowyn Tree AI

Description

Create a working model of the Elowyn Tree AI (LLMs) that demonstrates the practical applications of evolutionary AI in gaming and how it can guide the evolution of the EM of the Moloch AI.

Deliverables

Functional code with a tested working model of the Elowy Tree AI (LLMs) that showcases the practical applications of evolutionary AI in gaming. Ensure that the EM of the Elowyn Tree AI can be readily implemented in the Hyperon core architecture. Time estimate: 2 months.

Budget

$5,000 USD

Success Criterion

The EM of the Elowyn Tree AI is functioning and can be tested compared to traditional LLM output evaluation and context management through an open-source implementation of evolutionary DNNs, hosted on a public repository with an appropriate OSS license.

Milestone 4 - Evaluation, Testing & Technical Report

Description

Create a comprehensive guide and technical report that explains the experiments so they can be replicated for other AI approaches and include a summary of the evaluation and testing of the EMs that are developed through the project.

Deliverables

1. Provide a comprehensive guide for replicating experiments modifying DNNs and implementing other AI approaches including research results of evaluation and testing Time estimate: 1 month. 2. Provide a Technical Report: Detailed analysis of the evolutionary methods used including strengths weaknesses and comparative results with standard DNNs. Time estimate: 1 month.

Budget

$5,000 USD

Success Criterion

A comprehensive research guide and technical report have been created, with a focus on replicability and integration within the core architecture of Hyperon. In particular, the usefulness of applying the genetic algorithm to a massive DNN, using a multi-model approach.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

Group Expert Rating (Final)

Overall

3.0

  • Feasibility 2.8
  • Desirabilty 3.3
  • Usefulness 3.5
  • Expert Review 1

    Overall

    2.0

    • Compliance with RFP requirements 1.0
    • Solution details and team expertise 2.0
    • Value for money 3.0
    Detached from the technical challenges

    The proposal fails to address the RFP's exploration into CMA-ES. The high-level goals (game integration, behavioral testing) lack technical meaning and connection to the low-level building blocks (ES, hyperon integration).

  • Expert Review 2

    Overall

    3.0

    • Compliance with RFP requirements 3.0
    • Solution details and team expertise 3.0
    • Value for money 3.0
    Complex scope

    Highly creative and aligns with SNET’s mission. However, it needs sharper technical focus and stronger integration with Hyperon to meet the RFP’s primary goals effectively.

  • Expert Review 3

    Overall

    2.0

    • Compliance with RFP requirements 2.0
    • Solution details and team expertise 4.0
    • Value for money 3.0
    Interesting proposal but it's mainly about integrating Hyperon with a game-world, more-so that evolutionary DNNs

    It seems that after all the work needed for integration w/ the game world and dealing with that particular application, there won't be much time / resource left for the Evolutionary DNN work that is supposed to be the crux here...

  • Expert Review 4

    Overall

    5.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 4.0
    • Value for money 5.0

    Innovative and creative use of EM through creation of “an adaptive evolutionary learning environment” within the context of gameplay, all within DNNs.

feedback_icon