Does it ever make sense to abandon objectives in order to achieve those objectives? Perhaps counterintuitively, the answer is sometimes “yes!”
It was early morning when I got a weekly blog from my business coach. As I was reading the first few lines, I could clearly hear my mind was whispering “no” somewhere in the background. You know what I am talking about… That tiny person that likes to provide commentary when we try to listen fully or to read something that’s supposed to be interesting. Anyhow, it seemed like I was disagreeing with my coach. The subject of the article was “goals”, “intentions”, and how to accomplish them via simple daily steps as opposed to taking a long time to develop a strategy while waiting and doing nothing effective. The base premise was that to achieve our goals, we need to act, making small steps and reevaluating our progress as we go. We never have all the facts initially and as we move forward more information will help steer us in the right direction. Waiting to see the path delays us to no avail.
Should we try to make progress and reevaluate with every step?
Let’s think about it for a moment, we are stepping into a maze with high walls. We will keep moving forward (hitting walls, getting lost, etc.) and keep evaluating how far we are from our goal, but will this get us closer to a solution or get us there faster?
In the case of a maze, the answer is no. A large enough maze can confound us for a very long time. Further, most problems in life that are worth solving are not as simple as a maze.
In machine learning, we like to define a “fitness” function as an expression of progress towards an objective in the search space. Effectively we force the algorithm to act as an objective function. A good example (that works) will be a simple maze and a robot that is choosing a direction based on how close it gets it to the goal. Through deception, such objective functions may prevent the objective from being reached. See the maze on the right. We enter at #1, our goal is #3, and we get stuck at position #2. While trying to reach the center we will always get stuck in a location that is close, but any further move requires us to move away from the center which fails the “fitness” function and cause us to stop.
While methods exist to mitigate deception, they leave the underlying pathology untreated: “Objective functions themselves may actively misdirect search towards dead ends.”
Which means that I need a better plan or a strategy.
We need to solve a maze of goals, going from point A to point B knowing that without a good strategy we may never achieve our goals.
In general, for our discussion here, we will assume that ‘problems worth solving’ are seldom simple nor straightforward, and most simple-looking-problems are most likely deceptive. Not realizing a deceptive problem for what it is can be costly.
The second question that comes to mind is
Can we be spiritual and achieve our goals at the same time?
In Eastern spirituality, we learn to avoid attachment to objectives. While it sounds simple, by itself it is an objective, thus creating a circular dependency. As we observe people from the outset, to “abandon objectives” seems to be very confusing, random behavior and feels like there is no control over achieving our goals. It is challenging, for us humans, to comprehend random. In machine learning, this may be simpler. In machine learning we are not personally invested, thus if a goal is not achieved we are ok. This makes it simpler to take risks and try something new. Abandon objectives and see what happens.
Mathematically speaking: To abandon objectives, we need to simply abandon the use of the goal in our fitness function.
How does this work?
Let’s go back to the maze issue:
As human beings, we normally adopt the idea to go places we have never been before. We can ask a robot to do the same, although the goal is to reach the center, we will ask the robot to continue and move but always to a place which it has never been before. This is a strategy that does not include the objective as a directive, yet it will cause the robot to reach the goal at one point. One may think that this way it takes a long time to reach the goal, and this may be true, the goal will be reached.
To improve the above strategy, some machine learning deploys multiple agents/robots at once and requires each robot to be unique, e.g. always go to where other robots did not. This allows for a much faster convergence on a correct solution, and it does not include the goal as part of the heuristic. When we abandon objectives, we can still have a strategy or heuristic that leads to a solution. Defining this heuristic is sometimes very complicated and may require time to figure out.
A trickier example is the use of networking to achieve sales goals. The straightforward idea for a salesperson is to network towards leads which in turn help him/her to reach their sales goals. They network to identify potential clients. In this, the salesperson is what we call a ‘go-getter’. Some organizations choose to be ‘go-givers’. Those organizations are using a different heuristic. Instead of networking to achieve the business goals, the networking is centered on referring other people and creating business opportunities for others. This approach is very different. It does not include the objective as part of the heuristic but at the same time strengthen relationships and leads to new business creation.
To abandon objectives doesn’t necessarily imply that our goals will not be reached, it only forces us to have a heuristic approach that defines how we are making every small step along the way.
For myself, I rather choose a heuristic to represent ‘who I am’ and consider ‘what I leave behind when it is all said and done’.
Something in me says yes to this approach. Any other thoughts are welcomed.
⋅ ⋅ ⋅
 Abandoning Objectives: Evolution through the Search for Novelty Alone, Joel Lehman and Kenneth O. Stanley, Evolutionary Computation journal, (19):2, pages 189-223, Cambridge, MA: MIT Press, 2011
This post originally appeared on the author’s blog on multinnovation.com.