Throughout this series on Artificial General Intelligence, one term keeps popping up without much in the way of an explanation, the Workflow. What exactly is a workflow and what is its role in an AGI?
A workflow is a high-level description of a series of steps and sequences to be performed in order to achieve a pre-defined outcome. Consider the following diagram:
The above diagram describes a series of discreet activities which must be performed to ‘Update Weather’. Where we encounter branching, these activities can be performed in parallel. Each step must be completed in sequence, from ‘Start’ to ‘finish’ for the task to be fully complete. A particular branch can fail without impacting other branches, but the task would be marked as partially complete.
In an AGI, workflows underpin the bulk of intelligent activity by the platform. To be a General Intelligence, an AGI should be able to process any valid workflow and the workflow is the first step in approaching any assigned task. The workflow describes the ‘how’ of the approach to any particular task.
Each of the above discreet activities map to one or more programs which govern the particular implementation of the high-level description. For example, the high-level description of ‘Obtain weather information for all supported regions’ may map to a program which calls the APIs of regional/national weather providers and packages that data up in a database table.
A full blown AGI may have millions of workflows, which map to hundreds of millions of programs to perform a vast array of tasks. Mid-range, or expert, AGIs may only have a certain subset of that whole and are limited as a result.
For example, Ford may have a mid-range AGI which contains all the necessary workflows to operate as a mechanic to their vehicles, but it knows nothing of cooking, clock making, photography, etc. It could certainly learn, but that is not its role. In this case, this AGI hosted by Ford would be an expert AGI which a full blown AGI may make calls to, or defer to, when faced with the problem of resolving a malfunction.
The sheer quantity of programs to be produced is why an AGI is taught to program either as a first order approach, or very early in its development lifecycle. It would not be economically viable to have humans substantially program an AGI.
The workflows are a human readable description of the activity of an underlying HPC/Supercomputing solution. An orchestrator which spins up programs on-demand to perform the discreet activity. This could be a simple serverless function, or a few thousand servers to render a movie or use AlphaGo. It all depends on the task, priority, resources available, etc.
The workflow approach is also highly secure as certain discreete activities can be mandated on a universal basis. For example, if a workflow employs the use of a robotic arm, policies would automatically dictate that additional workflows be loaded which safe guard users in the vicinity. As such, there is a consistancy regardless of the specifics of the task.
In making an AGI intelligent, part of the process is teaching it how to create workflows and to leverage its vast program database to solve never before seen problems. While a large database of workflows will be sufficient for many problems, there will be scenarios where they need to be re-worked, often on the fly, to meet the constraints of a unique variant of a given task. This, obviously, is where intelligence comes into play but it is also guided and supported by the workflow engine, particularly in areas of policy, be that security, political requirements, etc.
DeepThought News 