Decisions about the provision of community services and infrastructure often involve multiple stakeholders with heterogeneous preferences. We formulate these problems as multi-agent, multi-constraint project portfolio selection tasks with private stakeholder utilities and propose an evolutionary negotiation mechanism for participatory coordination over feasible portfolios. In this approach, stakeholders are represented by software agents that negotiate over candidate solutions within a population-based search process. Proposal generation relies either on classical evolutionary variation operators or on learning-based strategies. Specifically, we iteratively retrain autoencoders and variational autoencoders on the evolving solution population to learn latent structural patterns induced by stakeholder preferences and resource constraints. New proposals are generated by recombining latent codes and decoding the resulting representations into candidate solutions. Computational experiments indicate that the learning-based proposal strategies produce solutions near the Pareto front, with variational autoencoders achieving the best performance in the studied scenarios.