Abstarct: Real-life decisions are usually made under conditions of uncertainty. In this thesis, definitions of the space of indeterminate variables are presented. In the following, the concept of EVaR is extended as a new uncertain coherent risk measure. Using the Chernoff inequality for VaR, we represent it as the lowest upper bound that can be found. Then, two types of loss function risk measures for uncertain systems are presented. First, the concept of VaR of the loss function is introduced baxsed on the uncertainty theory and its fundamental properties are examined. Then, the concept of TVaR of the evolved loss function sequence and some basic properties of TVaR are examined. Then, a combined portfolio optimization problem with mature securities and newly accepted securities are considered. Random-uncertain variables are used to describe the returns of securities, and to measurement of the corresponding risk, the related TVaR is introduced. Several mean − TVaR models are then formulated for the combined portfolio optimization problem, and the deterministic equivalents of these models are stated. As the main work, in this thesis, we deals a portfolio optimization problem with uncertain returns, where the returns of risky assets are as uncertain variables and are estimated by experienced experts. For this purpose, we first give a new mean-conditional value at risk-entropy model for the indeterminate portfolio optimization problem by considering four criteria, includes of return, risk, liquidity risk return and the degree of portfolio diversification. In the suggested model, as a bi-objective optimization problem, the investment return with an uncertain expected value and the investment risk with a conditional risk value are specified. Also, Entropy is used to measurement the degree of portfolio diversification. Then the uncertain models in different states are converted into equivalent deterministic optimization problems. Utilizing the weighted sum method, the bi-objective problems are reduced to single-objective problems. A capable neural network with reduced complexity is constructed for solving single objective problems. Finally, we provide several numerical simulations to confirm the efficiency and applicability of the presented scheme.