A systematic, multidimensional search of the space of theories of sentience, using simulated worlds and a genetic algorithm.
Why search systematically?
When you lose something at home, you first look in the logical places: your pockets, the table, the coat by the door. But when the situation becomes desperate, and you know for sure that the object must be somewhere in the house, you change strategy. You start to think of the house as a three-dimensional space, take into account the size of the object, and search it systematically, even in places that do not seem logical, until the phone turns up in the fridge or the ring in the sink.
Theories of sentience have developed in the same way. The first ones we consider are the most intuitive, or the most popular in our culture or our scientific community. That is a reasonable place to start. But we can also search the whole space of possibilities systematically, including theories that do not seem intuitive at first.
This matters for a practical reason. Popularity is an important factor when we assess a theory, but it is not the only one. If our goal is to reduce suffering, we should also pay attention to theories that may be as probable and reasonable as the popular ones, and take into account what would follow if they were true. A theory that few people hold, but which implies that a great deal of intense suffering is taking place, deserves a place in a plan to prevent suffering. A systematic search can help us find such theories.
The idea
The project creates a world in which the agents are not organisms but metaphysical hypotheses about sentience, and lets them evolve. In other words, it is a genetic algorithm that explores the space of possible theories of sentience.
- Each agent is a theory, with all its parameters: who can feel, what can be felt, how much, when, where, and what sentience is made of.
- The space of possibilities is defined by maps of theories of sentience. Before the search can start, these maps need to be developed so that they include every relevant dimension. Otherwise the algorithm can only find theories inside the region we already know.
- The fitness function compares the outcomes or predictions of each theory with reality.
The fitness function
All the difficulty of this project lies in the fitness function. To evaluate an agent, a simulation is run with the parameters of that theory, and the theory is scored according to two criteria:
- Coherence with the real world. Does the simulated world produce the kinds of beings, behaviours and reports that we actually observe?
- Internal contradictions. Can the theory be represented at all without its rules contradicting each other? A theory whose world falls apart is penalised.
An example run
- 100 theories of sentience are created at random, with random parameters taken from the space of possibilities.
- They are evaluated one by one. 50 are discarded (they “die”) because they are absurd.
- The coherence of the remaining theories is scored from 0 to 100. None scores more than 12 points.
- The surviving theories reproduce, combining their parameters.
- Mutations are introduced: random changes in the offspring.
- The population is shuffled.
- All theories that have not been evaluated yet are evaluated.
- The cycle is repeated, for example 10,000 times.
The low initial scores are expected. The point of the algorithm is not to start with good theories, but to see which regions of the space of possibilities improve over the generations.
A toy world with “supernatural” particles
To make the idea concrete, imagine a computer simulation that tries to be realistic. It contains material particles “floating” in their world, forming a kind of primordial soup. The particles combine into increasingly complex structures until, as the theory of evolution proposes, they form individuals like us.
In this simulation we can add supernatural particles: particles that do not necessarily follow the laws of physics. They are metaphysical particles.
I am not saying that these particles exist. What I am saying is that assuming their existence and representing them in a simulation lets us illustrate, understand better, and therefore assess a large number of theories about the capacity to feel.
The image shows four types of material particle, A, T, C and G, which combine according to certain rules. There is also a supernatural particle, shown in red, which could be an experience of pleasure or of pain. The supernatural particle also combines with matter according to a rule: it only “fits” next to three A particles in a row.
We can define the supernatural particles with complete freedom. Several of them could be in the same place at the same time, one could be in several places at once, or they could be available in unlimited amounts. They do not need to follow the laws of physics as we know them.
One world, many theories
With these few elements, different configurations of the same simulation represent different philosophical paradigms about reality and sentience:
- Emergentist monism. An individual acquires an experience when a certain material configuration (AAA) appears. We can then choose between two versions. In one, the experience has no effect on the material particles: it simply happens (epiphenomenalism). In the other, the experience changes the behaviour of the material particles.
- Immersionist monism. Symmetrically, we can treat the particles of pleasure and pain as the real ones and the material particles as epiphenomena. Individuals are then not generators of experiences but receivers of experiences, like a radio receiver that picks up a signal that already exists.
- Dualism. There are indeed two kinds of particle, material and experiential, and both are real.
Curiously, some configurations of this simulation make hypotheses about sentience compatible that previously seemed completely incompatible. When two theories produce the same simulated world, the disagreement between them may be about words rather than about facts.
Open questions
- How to build a map of theories complete enough to define the search space.
- How to measure “coherence with the real world” without favouring the theories we already find intuitive.
- How to prevent vague theories, which can accommodate any observation, from scoring too well.
- How to combine the score of a theory with its implications for suffering, so that improbable but very important theories are not simply discarded.
Related
- A map of the metaphysical theories of sentience
- Simulation of philosophical hypotheses about sentience
- This project is being developed with GaiaBasic, an open-source Python library for simulating theories of sentience.
Recent Comments