Could a Machine Fix Errors in its own Rationality?
Therefore, even if such a model could be created and executed, any predictions it makes would be useless, as they could not be communicated back to the physical world where our robot Plato is trying to determine whether to lead with a high trump card, or to play the King of Clubs.
A further nail in the coffin of the perfect rationality is the assumption of a deterministic universe; that the universe in a given state will necessarily progress to one single other state. However, modern physics places an inherent unpredictability in the physical world, embodied by the Heisenberg Uncertainty Principle.
4. All Errors in Rationality Can be Determined through Reflection and Peer Feedback
It’s simply not possible for any entity to have an accurate and subjective view of its own behaviour in the world, and thus cannot perceive flaws in that behaviour with complete confidence. To have any hope of improving its rationality beyond minor refinements, it must be able to evaluate its behaviour from an impartial ‘third-person’ perspective. This perspective can be gained from feedback from its peers. In the simplest case, imagine Plato having jumped out of the window, and survived, if a little shocked and shaken. It records the fact it has fallen, and decides that next time it won’t choose the brightest rectangle to leave by. This information may be useful if it ever finds itself in a very similar room, but wouldn’t really be reliable. Some doors are glass, and windows can have curtains for instance. Plato chances upon its friend, Socrates, and relates the event. As an older and more experienced robot, Socrates tells it that in future it should try to open the door and look before going through it.
Plato gains Socrates’ knowledge, and Socrates gains because its own theories have been confirmed by Plato’s fall. Combined with the feedback with guessing strategy of 2, such exchanges of information can save both robots a great deal of time and frustrated problem-solving attempts.
However, such feedback is far from reliable. Firstly, Socrates’ own views have been formulated by its own imperfect rational subsystem. Secondly, the feedback Socrates gives must be processed by Plato’s rationality. Thirdly, Socrates’ impartiality cannot be assured, particularly if the two robots are in competition for certain resources. So while information gained from peers has a potentially high value, it is no panacea for a faulty rationality.
5. Only By Becoming Irrational Could a Rational Machine Perceive its Flaws
The rationality dictates how an intelligence perceives and interacts with the world, and operates by constructing an internal model of the external reality. If the model is a faulty but overlapping (compatible) metaphor of the world, then a better model exists in the space outside the current model. By deliberately operating in this space, an intelligence must act in direct contradiction to its current rationality.
To detect flaws in its logic, a rational system must from time to time turn selected conclusions provided by the rationality on their head, and deliberately do the opposite of what its rationality advises.
The rationality’s purpose is to guide the intelligent machine to successful resolutions of problems it faces, yet fault-discovery involves taking action which is believed to lead to failure. In order for the system to become more successful, it must accept short-term losses. This strategy would appear to offer greatest gains to systems with especially flawed rationalities, as it is more likely that irrational behaviour will be successful.
A big problem is that a truly faulty rationality cannot reliably gauge risks it faces. Ideally, we would like our robot to determine when its existence is in danger and so when not to use the logic-inversion strategy, yet its ability to do so is limited when its rationality is very flawed.