thehumanbox
A browser-based civilization sandbox simulation project in which each simulated person is an independent tabular Q-learner. Newborns inherit their mother's Q-table along with mutation, and the project is written in Rust and compiled to WebAssembly.
Timeline (last 90 days)
Each simulated person in the "thehumanbox" project is an independent tabular Q-learner.
A newborn inherits a copy of the mother's Q-table, in which about 20% of the values are slightly adjusted (±0.03).
The child's traits are mixed and mutated from both parents, and it also inherits the mother's vocabulary and her attitudes toward other tribes.
thehumanbox is written in Rust and compiled to WebAssembly, running entirely in the browser with no backend required.
thehumanbox's action space has about 5000 possible actions, limited by era, age, location and the technologies the tribe has discovered.
In thehumanbox, action selection prioritizes survival needs and then executes an ε-greedy policy based on the learned values.
thehumanbox's relevant code is located in the open-source repository (MIT license) at crates/sim-core/src/sim/agents/growth.rs and organism/organism/learning.rs.
The developer candidly states that thehumanbox's two main weaknesses are that the state space is hand-built strings and the reward function is also hand-shaped.
The developer explicitly stated that he is curious about the RL community's views on thehumanbox's state space and reward design.
The design intent of the thehumanbox project is to explore the feasibility of "learning results as heritable units", treating the Q-table as a kind of "behavioral genome".
All Facts (10)
Each simulated person in the "thehumanbox" project is an independent tabular Q-learner.
50%UnverifiedA newborn inherits a copy of the mother's Q-table, in which about 20% of the values are slightly adjusted (±0.03).
50%UnverifiedThe child's traits are mixed and mutated from both parents, and it also inherits the mother's vocabulary and her attitudes toward other tribes.
50%Unverifiedthehumanbox is written in Rust and compiled to WebAssembly, running entirely in the browser with no backend required.
50%Unverifiedthehumanbox's action space has about 5000 possible actions, limited by era, age, location and the technologies the tribe has discovered.
50%UnverifiedIn thehumanbox, action selection prioritizes survival needs and then executes an ε-greedy policy based on the learned values.
50%Unverifiedthehumanbox's relevant code is located in the open-source repository (MIT license) at crates/sim-core/src/sim/agents/growth.rs and organism/organism/learning.rs.
50%UnverifiedThe developer candidly states that thehumanbox's two main weaknesses are that the state space is hand-built strings and the reward function is also hand-shaped.
50%UnverifiedThe developer explicitly stated that he is curious about the RL community's views on thehumanbox's state space and reward design.
50%UnverifiedThe design intent of the thehumanbox project is to explore the feasibility of "learning results as heritable units", treating the Q-table as a kind of "behavioral genome".
50%