A live demonstration and high level overview of this project can be found here. Stay tuned for a more in depth litepaper explaining the methodology in detail.
DeepBrew is an open-source research project which began as an undergraduate thesis topic and has expanded into an area of focus for Deeplink. It is an on-chain implementation of The Beer Game, a famous macroeconomics board game in which four players must optimize a supply chain, exchanging beer for money-our version uses ERC20 tokens on the Goerli Ethereum test-network, on which players are represented by smart contracts, and are given algorithmic behavioural strategies inspired by the popular 'base-stock' policy for The Beer Game. This game is also then represented as an OpenAI Gym class for reinforcement learning.
In addition to the game environemnt, DeepBrew also showcases a soft actor-critic deep Q-learning model tuned to play the game optimally as the distributor agent. This research project effectively proposes and demonstrates a prototype methodology for the recreation of a Web3 problem as a reinforcement learning problem, the training of an AI agent to optimize that problem, and the connection of these components to EVM blockchains, an amalgamation which we refer to as 'on-chain agents'. The lessons learned from the development and deployment of this system will be utilized in Deeplink's upcoming 'Cluster' projects.
We propose that DeepBrew's On-chain Beer Game can serve both as a demonstration of Deeplink's proposed 'Cluster' architecture, and as a standard for evaulating the performance of off-to-on-chain machine learning techniques. We also would like to invite the broader community to try their hand at further optimising and improving the system with their own machine learning and middleware solutions.