Second application to Vicarious
See the first application for context.
https://medium.com/@emenoh/i-applied-to-vicarious-8b4b6bae8dcb
Let’s do this again.
The general question is “How I would build intelligent machines”
The specific direction is: “Write a concise technical summary (500 words or less) explaining how you would go about building intelligent machines.”
Submission
Gathering of knowledge: An agent program would accept input (filtered/processed percepts — see below**) and compare with prior data using locality sensitive hashing (LSH) to find matching concepts in a Graph DB structured to map multiple sources to standardized label/graph subsets. Each subset of the graph is a property graph of an element of the agent world model (e.g. agent=> visual=> field=> object=> properties). The model includes at least: sensory input, objects and other agents, events and environmental states, goals and actions, costs and rewards. Responses to input are trained/pre-programmed, learned or a combination. Response feedback includes the input parameters as part of the return value and is labeled appropriately but is otherwise treated the same as knowledge gathering.
Actions (and reactions): Actions are easy, just do stuff. Actions can be modified to include pre-observations however. This is an important step and much harder. Preparing for an action or reaction is a learned behavior. Cost penalties (goal planning see below*) for discovering a state change in the environment abruptly should be appropriately high. The environmental model must include probabilities for abrupt state changes. When the agent seeks to transition to a scenario with high probability of abrupt change, the cost evaluation will prompt it to pre-observe (e.g. approximation of caution, fear, danger). The agent can learn about new dangerous scenarios by observing the frequency of state changes and assigning cost.
†Goal/Plan setting: a derivative of Goal Oriented Action Planning (GOAP) with a multi-step resolver and weighted costs system. e.g. proposed desired end-state(s) and map out possible actions to achieve (direct costs), changes to the environment (indirect costs), external events (indirect costs) and future freedom of action modeling (future costs). Complex goals of more than X steps will be broken into smaller goals with a re-evaluation step in-between. State changes also force re-evaluation.
††Percepts are filtered as streams using a set of parallel pipelines of transformations on sampled data. The return of each pipe will attach properties to the input that can be used to perform the LSH query above against the Graph DB store.
A portion of the graph will be used to model the current environment in a persistent way but will be copied and flattened into a key/value list/dictionary to provide shortcuts for reflexes if the agent is autonomous (e.g. daytime?, alone?, raining?). New shortcuts will be learned using a frequency of status / discreteness of value sampler. If an LSH query returns the same set of discrete values repeatedly and the change in state is infrequent, a new shortcut/assumption will be created. Assumptions will be tested randomly to verify that a potential value has not been missed.
Specific transformations on data-types can be discussed at length in an interview — gradient descent/bayesian models apply but application is context specific; abstracting out useful data points becomes paramount. Other areas left out: internal model, cost/reward model (focused on utility in actions), reaction triggers.
Thanks! Unlikely I’ll hear from you but was fun to write.
A few references
