Computational Progress Report

    In this essay, I will summarize the computational strategy that incorporates Dr. Brandt’s blending model dealing with intention into the current algorithms and capacities of the agent. This general plan revolves around the two different actions the agent currently has, namely explore and eat. The architecture is devoted to weighing the options of each, selecting one, simulating the action, executing it step by step, and evaluating the progress and relevancy of the action at each step.

    First, the model begins with the state the agent is currently in. This corresponds to the pragmatic space in the blending model, with the only difference being that there is no human requesting an action from the agent because the program is not yet sophisticated enough to support such operations. Instead of human interaction, the agent basically has a fuel or energy meter at this point. The format of this meter is not terribly important, most likely a scale of one to a hundred, with the only important aspect being certain ranges influence how the agent acts. For example, if energy is below 25%, then abandon all current objectives and look for food, or if it is about 75%, then stop looking for food and just hang out or go explore.

    The decision is calculated based on a choice between different objectives. The current objectives are eating or exploring, and there are multiple instances of each because there are different things to eat and multiple avenues to explore. Each other these possibilities have a certain cost associated with it. This cost takes into account the distance of the goal and therefore the amount of fuel it will use, and whether or not the trip has a projected award (like an energy supplement). Calculating the cost is the goal of planning.

    During planning, the type of motion involved is defined, in this case either moving or looking (in order to eat, the agent just looks at luminescent objects). The objective resides in the presentation space of the blending model, corresponding to a general algorithm that will be simulated in order to ascertain the effect of the actual action.  In order to properly simulate this general goal, the algorithm needs the absolute map as an input. This corresponds to the blending in the model, making the generic action specific by mapping it onto the third person representation of the environment. This simulation creates a stepwise action plan with the steps being landmark points of intended evaluation.

 From this simulation, the agent knows how much fuel it will take to get there, the predicted amount of fuel it will receive, taking into account the traversibility of the path (whether the regions are smooth or rough, if there is a wall) and it will calculate a projected ‘cost’ of deciding to do this action. Each objective will have a cost like this in a kind of ‘network of blends.’ For example, there may be many different potential paths of exploration and a few known energy sources on the map. Each of them will be simulated and assigned a cost based the distance and reward, and depending on the robot’s current state, it will choose the least expensive option. Or in the case of starvation, it might be cheap to go explore close, but the gas tank forces the agent to go to a known refuel point rather than explore that other territory.

    Once a decision is reached based on the cost function, the agent acts out the first step of the sequence of events. Next, the environmental input, and therefore the cost of different objectives, changes. The agent updates its blending network to accommodate for these changes and evaluates whether the current action is still worth doing, or decide another action is better suited to fulfill its desire.