The evolution of goals in AI agents

摘要

Forced evolution has been proposed as a possible path to developing artificial general intelligence. For practical reasons, self-replicating robots are being proposed for missions where direct manufacture could be prohibitive or as a cost-effective means to maintain a stable working population of robots. If self-replication occurs in a harsh (i.e. selective) environment, the forces of evolution may distort the originally programmed objectives. Via millions of simulations of AI agents with nematode-level neural networks, this research explores the consequences of allowing replication in a hostile and competitive environment. As the selection pressures are tuned, the evolution of their neural networks and corresponding behavioral changes are tracked. As a consequence of these simulations, agents with multi-layer neural networks trained simply to retrieve resources, consume needed resources, and evade obstacles evolve behaviors that look like evasion of hostile overseers, the intended murder of enemies, and cannibalism of other agents. These simulations are intended to directly address safety concerns around creating self-replicating AI agents or robots. As designers, if we allow replication under selection pressure, regardless of initial designs, we risk allowing the emergence of unintended strategies. One solution to preventing evolution could be to enable AI agents with continuous backup– immortality.