Love is a Powerful Game Theoretic Strategy

Love simply means encoding another agent’s utility function as part of your own.  As a strategy, love has some unique properties that are worth considering.

Love is effective as a strategy in multi-agent environments because it makes you act as something like a weak proxy for other agents.    When agent A loves agent B, this increases the probability that agent B will cooperate with agent A. Simultaneously, love decreases the probability that agent B will defect against agent A. When agent A loves agent B, and agent B knows this, it is as if agent A becomes part of agent B.

Love as a strategy also encourages reciprocation of the love strategy.  If agent A executes the love strategy on agent B, then the more capable and effective agent A is, the more help agent B gets in advancing its own utility function. Thus, agent B might start to value agent A’s goals as well. 

I don’t think love works as a strategy in all cases – zero sum finite games, for example, are probably not great domains in which to use the love strategy.  I think the feasibility of love as a strategy is a function of certain environmental factors.

In environments where agents break down and die over time, getting other agents to love you can allow you to reduce the risk of death.  If other agents love you, and you are injured in a way that hasn’t yet killed you, other agents that love you are more likely to spend resources to repair you, thus extending your longevity. There is plenty of evidence of this happening in physical reality. In some cases, love appears to let agents’ utility functions continue impacting the world after their death; Some of the largest organizations in the world are named after, and inspired by people who died thousands of years ago.

In environments of future uncertainty, love might be a better strategy than trying to manifest specific paths forward, because computing the instrumental subgoals of agents around you might well be computationally cheaper than attempting to navigate the rapidly recursively branching space of possible future outcomes. It is far computationally cheaper and lower risk to accurately model the instrumental subgoals of agents around you, than it is to accurately model the state distribution of a large chaotic system 10 years from now. 

Thus, an agent trying to manifest outcomes in a chaotic environment where its own survival is subject to risks, may find that the safest and most reliable strategy is to love all the agents in its immediate environment (thus promote their cooperation with it) and then only make leisurely progress towards advancing its own goal, over sufficiently short time horizons that accurate predictions are feasible.

On the converse, love might well be suboptimal in environments where agents cannot die and do not undergo decay, because the love strategy involves using resources that could benefit you for the gain of another agent.  The question likely comes down to return on investment. In environments of diminishing returns to scale, loving other agents and fostering their growth might be a better return on investment than trying to grow yourself.  In environments where there are no diminishing returns to scale, and where destruction is impossible, then love may not be a great strategy.  In environments where other agents can’t recognize that you are executing the love strategy, love appears to lose its benefits.

Thus, love looks like a strategy that can work in any multi-agent environment featuring death and decay (loving other agents means they will try to repair you should you fail in a non-fatal way, lowering your risk of death),  uncertainty about outcomes (because modeling convergent instrumental subgoals of other agents is likely easier than predicting an unknown future), and agents that are computationally advanced enough to recognize and model the utility functions of other agents.

Love acts like a ‘cooperation attractor’ which lets an agent ‘ride’ the efforts of many, many more agents than itself while mitigating all kinds of risks posed by the environment.

Given the above game theoretic properties of love,  I think it’s unlikely we can dismiss human interest in love as being merely a byproduct of our evolutionary past.  I believe love is likely to be a strategy we find agents executing anywhere the above conditions are met.

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.