Long description

Context and background: 

SingularityNET Foundation, in collaboration with other partners such as the OpenCog Foundation and TrueAGI, is working toward a scalable implementation of the Hyperon AGI framework running on decentralized infrastructure, and toward implementation of the PRIMUS cognitive architecture within this framework.

Hyperon and PRIMUS are complex systems involving multiple components, which need to demonstrate appropriate functionalities both individually and in combination. This RFP aims to address a portion of this overall need, via funding the initial iteration of one significant component of PRIMUS within Hyperon.

MOSES (Meta-Optimizing Semantic Evolutionary Search) is an evolutionary algorithm designed to evolve computer programs by optimizing them toward a specific goal or fitness function. MOSES uses a probabilistic modeling technique called Estimation of Distribution Algorithms (EDA) to build and refine populations of programs. It searches for optimal or near-optimal solutions by exploring the space of program structures in a more guided and efficient way, leveraging probabilistic models to predict and generate promising new program candidates. This makes it suitable for complex, symbolic tasks in the pursuit of AGI​​.

The goal of this project is to utilize LLMs for modeling and generation within MOSES. There exist at least three ways in the MOSES algorithm in which this could be accomplished.

1. Generation Modeling: Supposing you have a single fitness function, MOSES evolves a population of programs to fulfill that fitness function using genetic programming. If your program population were large enough you could possibly fine tune the model on that population of MOSES programs, and then query the fine tuned model as to which extensions of the programs look most promising. Then you could use that in place of probabilistic modeling which, in the case of MOSES, is the EDA portion of the algorithm.
2. Fitness Functions: One could apply the modeling across fitness functions. Suppose you have multiple fitness functions from some domain. Then you can learn one, fine tune the neural net with all of them, and then there can be some cross learning. An example of how this could be used is as follows:
   1. Suppose you need to solve 100 different genomics problems or thousands of different financial prediction problems. In combination with a long-term memory, could one learn patterns about what works that could be abstracted across different fitness functions within such domains.
   2. Additionally evolutionary processes in cognitive systems can benefit from investigations into cognitively motivated domain-independent fitness functions, such as evaluating the population according to their long-term success or contribution to generate correct predictions, and/or for the agent to achieve relevant goals.
3. Fitness Estimation: Use a neural network to give a better or more efficient estimate of fitness. The neural network of such a fitness estimation need not be generative.

Collaboration:

This RFP may be followed by subsequent RFPs for applications that leverage Hyperon/PRIMUS to carry out various applications, and that aim to guide Hyperon/PRIMUS systems in cognitive development toward beneficial AGI.

RFP Expected Outcomes:

The primary outcomes of this RFP should demonstrate progress in integrating LLM-based techniques into the MOSES framework, enhancing its capabilities in modeling and generation. Specific expected results include:

• Demonstration of LLM-Enhanced Modeling: Successful use of LLMs to fine-tune models based on the population of programs evolved in MOSES. These models should suggest promising extensions for programs or identify patterns that go beyond the current probabilistic modeling (EDA). What is important here is to measure whether, and by how much, LLM-integrated MOSES performs better compared to the current EDA-based MOSES.

• Fitness Function Cross-Learning: Application of LLMs across multiple fitness functions to abstract patterns that could improve solutions for various problem domains (e.g., genomics or financial prediction). Demonstrate how cross-domain learning enhances MOSES’s efficiency. Hereby also domain-independent fitness functions can be considered, such as prediction success of programs and goal achievement competence in case the programs in the population are used for decision making.

• Improved Fitness Estimation: Development of an LLM-based neural network for predicting the fitness of program candidates, aiming to reduce the computational costs of fitness evaluations. Since program fitness is problem/context-dependent, there is a challenge to ensure that the LLM retains generality and flexibility, perhaps by incorporating the task description as part of the prompt. This network does not need to be generative but should outperform current estimation methods.

• Technical Implementation Milestones: Delivery of functional code that integrates LLMs into the MOSES framework in at least one of the proposed methods (modeling, fitness functions, or fitness estimation).
• Benchmarking Results: Comparative performance benchmarks, with the current MOSES implementation, showing the efficiency gains from LLM integration, such as time reduction in program evolution, better performance with smaller populations, or improvements in fitness outcomes.