In this paper, we present DALI, a self-adaptive, collaborative multi-agent Traffic Signal Timing system (TST). Intersection controller agents collaborate with one another and adapt their timing plans based on the traffic conditions. Reinforcement learning is used to optimize values for the various thresholds necessary to dynamically determine the scope of collaboration between the agents. DALI was implement in MATISSE 3.0, a large-scale agent-based micro-simulator. Experimental results show an improvement over traditional and reinforcement learning TSTs.