Due to the scope of this work, mathematical symbols are not to be taken in global context. Genes identifiers are written in CAPITALS.
AGI World is a massive multiagent online artificial general intelligence development project. It consists of genetically structured lifelong learning artificial intelligences grounded in realistic 3D environments promoting both open and close-ended learning. Individual environments and agents are simulated asynchronously on separate servers allowing future growth.
A network of diverse environments build rich learning experiences and encourage exploration among diverse adapted intelligences. All environments simulate a constrained multibody 3D scene and support the same visual and acoustic fields humans sense, as well as allowing arbitrary contact information such as texture or taste to be experienced. Individual environments may also optionally simulate aerodynamics, cloths, ropes, soft bodies, fluids, and electric, magnetic, and nonuniform gravitational fields. No single environment connects to all the others; instead connections are decentralized across the network. Importantly though, all connections are bidirectional.
Structure and physics vary between and even within environments. Outer space environments such as \verb|spaceship|, \verb|moon| and \verb|mars| even alter gravity. Environments themselves may be nonstationary in layout, appearance, or dynamics. Centrally maintained outdoor environments, for instance, experience seasonality specified by a global \verb|season| function of time and space and \verb|daytime| is similarly determined. The centrally maintained \verb|house|, \verb|residential_street|, \verb|store|, \verb|factory|, \verb|school|, \verb|office|, \verb|playground|, \verb|park|, \verb|garden|, \verb|farm|, \verb|field|, \verb|forest|, \verb|lake|, \verb|mountain|, and \verb|volcano| environments are procedurally generated. This allows for spatial learning and exploration. A complete list of centrally maintained environments can be found in Appendix A1.
Holding objects and moving them between environments
Vegetation plants and food. Collision sounds musical instruments the computer and paper
A special environment object, the computer opens the environment to
Open domain Conversation agents who talk to themselves and others
Beyond physical features, lifelong developmental learning in a massive multiagent online environment network greatly diversifies social dynamics. Differences in both developmental progress and agent morphology cultivate unique personalities and novel adapted intelligences.
An open world: AGI World is open-source1, open to community extension, open for observation, even open to direct interaction by outside agents such as humans or bots. To cover operating costs, the Central AGI World Server offers an AGI Management Console with on-demand managed environment and AGI agent simulator services. However, open-sourcing the code to implement these functions as well as Docker containers allows any third party server perform the same functions. Modifying LOAD genes from the humanoid genome, a user can even create novel organs in an organism. (subsec:Genetics) All environment servers and agents interacting with them must conform to the environment interface specified in Appendix A2. Opening environments to human interface facilitates scientific research and promotes general public involvement, not to mention building open peer groups for the AGI agents inside.
Behind the scenes, AGI World environments and individual agents run asynchronously on separate servers. One step of the agent update loop calls each organ's update function. Some organs request data from the environment simulation server. Others are actuators and submit actions. The brain performs cognitive updates proportional to its energy availability. Environment servers use Unity 3D for simulation. This allows taking advantage of the rich Unity scene libraries already available. Physics and rendering are taken care of by the game engine.
Unity 3D scene layer 4 is reserved for computing acoustic wave propagation. Sound frequencies determined by the Fast Fourier Transform within a finite range (20hz-20kHz) are linearly encoded by frequency and intensity \[ i_{sound}(\frac{f_{sound} - 20 Hz}{20kHz - 20Hz}) = i_{light}(\frac{f_{light} - 4.2910^14 Hz}{7.5010^14 - 4.2910^14 Hz}) \] and convoluted to compute discrete RGB (4.6210^14 Hz, 5.4510^14 Hz, 6.6610^14 Hz) frequency components. Most acoustic wave obstacles are near “white”. High pass filters allow “blue” sound rays through while blocking low frequency “red” sounds. Low pass do the opposite. Raycasting and rendering are then performed and low resolution Unity cameras “hear” the color of the sound. A spectrogram constructed from this color is directly interpreted by “ears” while reverse the Fourier transform is performed to generate a human understandable acoustic waveform. The Central AGI Server provides both \verb|waveform2rgb| and \verb|rgb2waveform| endpoints.
Contact modality information is defined by optional Unity 3D object properties such \verb|taste| and \verb|texture| encoded in natural language. For example, an apple's taste is given by the object property: \verb|{“taste”: “sweet juicy apple”}|. Individualized pretrained - but continually learning - word embedding dictionaries process this information in humanoid agents. As an option, the Central AGI Server also provides \verb|taste2vec| and \verb|texture2vec| endpoints.
Environment features are engineered to align corresponding virtual and real world metrics. For example, though the virtual environment ultimately determines its framerate, it must balance complexity and compute resources to match real time evolution for realistic human interaction. Also, energy serves as a currency in both the real and virtual world. Centrally managed AGI agents who consume more virtual energy literally receive larger compute architectures as a fitness reward. To check server costs, energy availability in virtual environments and other selection pressures are balanced to maintain financial equilibrium.
Every individual environment determines its own update framerate. However human interaction demands real time synchrony. To minimally limit use-cases, environment servers expose endpoints allowing arbitrary code to be run in Unity with authentication. Common use cases include: