Demand response (DR) is a recent effort to improve efficiency of the electricity market and the stability of the power system. A successful DR implementation relies on both appropriate policy design and enabling technology. Real-time pricing (RTP) and time of use (TOU) have been identified as two important DR policies to motivate residential customers to participate in DR programs. An efficient residential DR model should implement heterogeneous residential load forecasting, multi-criteria optimization (e.g., objectives for individual homes, utilities and aggregations of them) and intelligent distributed algorithms to evaluate the complex and large-scale power systems. This thesis presents a multi-agent system (MAS) to evaluate optimal residential DR implementation in a distribution network, in which the main stakeholders are modeled by heterogeneous home agents (HA) and a retailer agent (RA). A heterogeneous load prediction model, a real-time electricity price model and three optimal load control models are developed to associate with the MAS. The load prediction model simulates the benchmark of individual and aggregated load profiles based on statistical information of how people use their appliances including electric vehicles (EV). Each HA has a unique load profile depending on its heterogeneous local configurations. The real-time price prediction model is defined as piecewise linear functions of power and the optimal coefficients are obtained from historical data of real-time loads and electricity prices via the norm approximation approach. The optimal load control models are developed based on dynamic pricing of RTP and TOU. An open-loop optimal load control model under RTP (OL-LCM-RTP) is formulated into a convex programming (CP) problem to minimize electricity payment and waiting time. A HA schedules the controllable loads based on its local information by solving the CP problem; therefore, it only requires a minimum of communication between the HA and the RA. This is greatly useful because the infrastructure for communication is still under development. In addition, the privacy of users is not sacrificed. Simulation results show that the peak-to-average power ratio (PAPR) and the standard deviation of the load profile, and electricity payments are reduced using the proposed mechanism. A close-loop optimal load control model under RTP (CL-LCM-RTP) is developed based on the OL-LCM-RTP by further incorporating feedbacks from RA. A HA solves the CP problem to schedule the controllable loads in a round process using the global load information. The process can be quickly converged in the second round; therefore, it requires limited efforts from the communication and the coordination. It is found that this model can significantly improve the quality of the optimization based on the simulation results. An optimal load control model under TOU (LCM-TOU) is modeled by a linear programming (LP) problem. The objective function is designed to find a trade-off among three factors: 1) the minimum electricity payment; 2) comfort levels with waiting time; and 3) to avoid peak demand rebound. We also evaluate the impacts of the participation levels of TOU programs. The simulation results show a reduced PAPR, standard deviation and the electricity payments from the HAs. The HA, with proposed optimal control mechanisms, can be embedded into a home energy management system (EMS) to make intelligent decisions on behalf of homeowners automatically responding to DR policies. The proposed agent system can be utilized to evaluate various strategies and emerging technologies that enable DR implementation.