How Can Governance of Attention and Memory Change Choice?

· ·

Three decision governance strategies are compared in terms of how they influence the ability of an agent to reach their goal in a simple problem. The strategies are:

  • No decision governance: The agent is assumed to choose randomly between options, and they can see all the options in the problem, i.e., they have perfect information at each step, but have incomplete information about how steps lead to their goal. The agent does not remember their choices, and each step is new.
  • Governance of memory:  Decision governance is introduced to ensure agent’s choices are logged and the agent is discouraged from repeating them, that is, governance compensates for the agent’s absence of memory (memory in the sense that the agent would remember if they were in the same place twice). 
  • Governance of attention: Decision governance is such that it directs the agent’s attention at every step to the option that is closest to their goal.

This text is part of the series on the design of decision governance. Decision Governance refers to values, principles, practices designed to improve the quality of decisions. Find all texts on decision governance here, including “What is Decision Governance?” here.

The key difference between the governance of memory and of attention, as outlined above, is that you can implement governance of memory even if you do not know the exact goal of the agent (and so cannot guide them to it). Instead, you can only implement governance of attention – at least in the sense this has in the present text – without knowing the goal they are aiming to achieve.

Let’s consider a simple simulation to show the differences between the three strategies. We have an agent that needs to move through a grid. The agent starts in the upper left corner of the grid and its goal is to reach the lower right corner of the grid. We then set the agent’s behavior as follows, for each case:

  • No decision governance: In any cell of the grid, the agent chooses randomly one of the adjacent cells, and does so until it reaches the goal. The agent does not know where the goal position is, relative to its current position – it will know that it achieved its goal only when it reaches the goal position.
  • Governance of memory: The agent chooses in the same way as in the no governance case, but we now introduce rules:
    • all steps of the agent are recorded, 
    • when the agent chooses, it needs to consult the log, and needs to choose among random adjacent cells (a.k.a. positions) which it did not already visit, unless it visited all, in which case it chooses the random one among all adjacent cells,
    • the agent is thus discouraged from revisiting the same position. 
  • Governance of attention: Again, the agent chooses in the same way as in the no governance case, but we introduce these rules: in each position, we draw the agent’s attention to the next position which is closest to the goal position, without telling the agent what that goal position is. The agent’s attention is not drawn to anything else in any given position in the grid. (There is nothing they encounter as they move in the grid, which would steer their attention away from how we are steering them.)

For 1000 runs of the simulation for each case, the following is the summary of the number of steps and standard deviation in each case:

  • No decision governance: 6808.3 steps on average, with standard deviation of 6189.9.
  • Governance of memory: 2234.7 steps on average, with standard deviation of 3190.6.
  • Governance of attention: 59.0 steps on average, with standard deviation of 0, the agent always takes the same number of steps.

Below are histograms of numbers of steps across 1000 runs of each case above.

The image below shows three sample runs for each case. In an image, if an agent visited the cell, then a circle is drawn in the cell, and the more times the agent visited that cell, the less transparent the color of the circle.

Code for simulations is available at github, here.